[X86] Canonicalize extract_vector_elt from vXi1 to always return MVT::i32.

This allows us to remove some isel patterns that allowed MVT::i8 result type.

llvm-svn: 320936

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 7ee8bbc..7135be3 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14568,6 +14568,13 @@ static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG,
     return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
   }
 
+  // Canonicalize result type to MVT::i32.
+  if (EltVT != MVT::i32) {
+    SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
+                                  Vec, Idx);
+    return DAG.getAnyExtOrTrunc(Extract, dl, EltVT);
+  }
+
   // If the kshift instructions of the correct width aren't natively supported
   // then we need to promote the vector to the native size to get the correct
   // zeroing behavior.

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 55cf71a..97c05a5 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -2773,12 +2773,6 @@ let Predicates = [HasAVX512] in {
 
     def : Pat<(maskVT (scalar_to_vector GR8:$src)),
               (COPY_TO_REGCLASS (INSERT_SUBREG (i32 (IMPLICIT_DEF)), GR8:$src, sub_8bit), maskRC)>;
-
-    def : Pat<(i8 (X86Vextract maskRC:$src, (iPTR 0))),
-              (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS maskRC:$src, GR32)), sub_8bit)>;
-
-    def : Pat<(i32 (anyext (i8 (X86Vextract maskRC:$src, (iPTR 0))))),
-              (COPY_TO_REGCLASS maskRC:$src, GR32)>;
   }
 
   defm : operation_gpr_mask_copy_lowering<VK1,  v1i1>;

diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll b/llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll
index 908dcd7..f3ca064 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll
@@ -3007,13 +3007,13 @@ define <4 x float> @test_mask_vextractf32x4(<4 x float> %b, <16 x float> %a, i8
 ; CHECK-NEXT:    vmovd %edi, %xmm2
 ; CHECK-NEXT:    kmovw %edi, %k0
 ; CHECK-NEXT:    kshiftrw $3, %k0, %k1
-; CHECK-NEXT:    kshiftrw $2, %k0, %k2
-; CHECK-NEXT:    kshiftrw $1, %k0, %k0
-; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    vpinsrb $4, %eax, %xmm2, %xmm2
-; CHECK-NEXT:    kmovw %k2, %eax
-; CHECK-NEXT:    vpinsrb $8, %eax, %xmm2, %xmm2
 ; CHECK-NEXT:    kmovw %k1, %eax
+; CHECK-NEXT:    kshiftrw $2, %k0, %k1
+; CHECK-NEXT:    kmovw %k1, %ecx
+; CHECK-NEXT:    kshiftrw $1, %k0, %k0
+; CHECK-NEXT:    kmovw %k0, %edx
+; CHECK-NEXT:    vpinsrb $4, %edx, %xmm2, %xmm2
+; CHECK-NEXT:    vpinsrb $8, %ecx, %xmm2, %xmm2
 ; CHECK-NEXT:    vpinsrb $12, %eax, %xmm2, %xmm2
 ; CHECK-NEXT:    vextractf32x4 $2, %zmm1, %xmm1
 ; CHECK-NEXT:    vpslld $31, %xmm2, %xmm2
@@ -3032,13 +3032,13 @@ define <4 x i64> @test_mask_vextracti64x4(<4 x i64> %b, <8 x i64> %a, i8 %mask)
 ; CHECK-NEXT:    vmovd %edi, %xmm2
 ; CHECK-NEXT:    kmovw %edi, %k0
 ; CHECK-NEXT:    kshiftrw $3, %k0, %k1
-; CHECK-NEXT:    kshiftrw $2, %k0, %k2
-; CHECK-NEXT:    kshiftrw $1, %k0, %k0
-; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    vpinsrb $4, %eax, %xmm2, %xmm2
-; CHECK-NEXT:    kmovw %k2, %eax
-; CHECK-NEXT:    vpinsrb $8, %eax, %xmm2, %xmm2
 ; CHECK-NEXT:    kmovw %k1, %eax
+; CHECK-NEXT:    kshiftrw $2, %k0, %k1
+; CHECK-NEXT:    kmovw %k1, %ecx
+; CHECK-NEXT:    kshiftrw $1, %k0, %k0
+; CHECK-NEXT:    kmovw %k0, %edx
+; CHECK-NEXT:    vpinsrb $4, %edx, %xmm2, %xmm2
+; CHECK-NEXT:    vpinsrb $8, %ecx, %xmm2, %xmm2
 ; CHECK-NEXT:    vpinsrb $12, %eax, %xmm2, %xmm2
 ; CHECK-NEXT:    vpslld $31, %xmm2, %xmm2
 ; CHECK-NEXT:    vpmovsxdq %xmm2, %ymm2
@@ -3056,13 +3056,13 @@ define <4 x i32> @test_maskz_vextracti32x4(<16 x i32> %a, i8 %mask) {
 ; CHECK-NEXT:    vmovd %edi, %xmm1
 ; CHECK-NEXT:    kmovw %edi, %k0
 ; CHECK-NEXT:    kshiftrw $3, %k0, %k1
-; CHECK-NEXT:    kshiftrw $2, %k0, %k2
-; CHECK-NEXT:    kshiftrw $1, %k0, %k0
-; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; CHECK-NEXT:    kmovw %k2, %eax
-; CHECK-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; CHECK-NEXT:    kmovw %k1, %eax
+; CHECK-NEXT:    kshiftrw $2, %k0, %k1
+; CHECK-NEXT:    kmovw %k1, %ecx
+; CHECK-NEXT:    kshiftrw $1, %k0, %k0
+; CHECK-NEXT:    kmovw %k0, %edx
+; CHECK-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; CHECK-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; CHECK-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; CHECK-NEXT:    vextracti32x4 $2, %zmm0, %xmm0
 ; CHECK-NEXT:    vpslld $31, %xmm1, %xmm1

diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics.ll b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
index 35a502b..5faa202 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
@@ -3305,17 +3305,17 @@ define i8@test_int_x86_avx512_mask_cmp_sd_all(<2 x double> %x0, <2 x double> %x1
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_sd_all:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    vcmplesd %xmm1, %xmm0, %k0
-; CHECK-NEXT:    kmovw %k0, %ecx
+; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    vcmpunordsd {sae}, %xmm1, %xmm0, %k0
-; CHECK-NEXT:    kmovw %k0, %edx
+; CHECK-NEXT:    kmovw %k0, %ecx
+; CHECK-NEXT:    orl %eax, %ecx
 ; CHECK-NEXT:    kmovw %edi, %k1
 ; CHECK-NEXT:    vcmpneqsd %xmm1, %xmm0, %k0 {%k1}
-; CHECK-NEXT:    kmovw %k0, %esi
+; CHECK-NEXT:    kmovw %k0, %edx
 ; CHECK-NEXT:    vcmpnltsd {sae}, %xmm1, %xmm0, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    orb %cl, %dl
-; CHECK-NEXT:    orb %sil, %al
-; CHECK-NEXT:    orb %dl, %al
+; CHECK-NEXT:    orl %edx, %eax
+; CHECK-NEXT:    orl %ecx, %eax
 ; CHECK-NEXT:    ## kill: def %al killed %al killed %eax
 ; CHECK-NEXT:    retq
 
@@ -3350,17 +3350,17 @@ define i8@test_int_x86_avx512_mask_cmp_ss_all(<4 x float> %x0, <4 x float> %x1,
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_ss_all:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    vcmpless %xmm1, %xmm0, %k0
-; CHECK-NEXT:    kmovw %k0, %ecx
+; CHECK-NEXT:    kmovw %k0, %eax
 ; CHECK-NEXT:    vcmpunordss {sae}, %xmm1, %xmm0, %k0
-; CHECK-NEXT:    kmovw %k0, %edx
+; CHECK-NEXT:    kmovw %k0, %ecx
+; CHECK-NEXT:    andl %eax, %ecx
 ; CHECK-NEXT:    kmovw %edi, %k1
 ; CHECK-NEXT:    vcmpneqss %xmm1, %xmm0, %k0 {%k1}
-; CHECK-NEXT:    kmovw %k0, %esi
+; CHECK-NEXT:    kmovw %k0, %edx
 ; CHECK-NEXT:    vcmpnltss {sae}, %xmm1, %xmm0, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    andb %cl, %dl
-; CHECK-NEXT:    andb %sil, %al
-; CHECK-NEXT:    andb %dl, %al
+; CHECK-NEXT:    andl %edx, %eax
+; CHECK-NEXT:    andl %ecx, %eax
 ; CHECK-NEXT:    ## kill: def %al killed %al killed %eax
 ; CHECK-NEXT:    retq
   %res1 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 2, i8 -1, i32 4)

diff --git a/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll b/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
index cbd23ae..9863a0a 100644 (file)
--- a/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
+++ b/llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll
@@ -714,32 +714,32 @@ define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -778,32 +778,32 @@ define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -845,32 +845,32 @@ define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -913,32 +913,32 @@ define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -981,34 +981,34 @@ define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -1050,34 +1050,34 @@ define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>*
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -1122,34 +1122,34 @@ define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -1195,34 +1195,34 @@ define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -2480,83 +2480,15 @@ define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k0
-; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
-; NoVLX-NEXT:    andl $1, %eax
-; NoVLX-NEXT:    kmovw %eax, %k1
-; NoVLX-NEXT:    kxorw %k0, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
-; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
-; NoVLX-NEXT:    korw %k1, %k2, %k1
-; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
-; NoVLX-NEXT:    kxorw %k0, %k2, %k0
-; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
-; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
-; NoVLX-NEXT:    kxorw %k1, %k0, %k0
 ; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
-; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
-; NoVLX-NEXT:    retq
-entry:
-  %0 = bitcast <2 x i64> %__a to <4 x i32>
-  %1 = bitcast <2 x i64> %__b to <4 x i32>
-  %2 = icmp eq <4 x i32> %0, %1
-  %3 = bitcast i8 %__u to <8 x i1>
-  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
-  %4 = and <4 x i1> %2, %extract.i
-  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
-  %6 = bitcast <8 x i1> %5 to i8
-  ret i8 %6
-}
-
-define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
-; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
-; VLX:       # %bb.0: # %entry
-; VLX-NEXT:    kmovd %edi, %k1
-; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovd %k0, %eax
-; VLX-NEXT:    # kill: def %al killed %al killed %eax
-; VLX-NEXT:    retq
-;
-; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
-; NoVLX:       # %bb.0: # %entry
-; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %edi, %k0
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -2592,63 +2524,131 @@ define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    retq
 entry:
   %0 = bitcast <2 x i64> %__a to <4 x i32>
-  %load = load <2 x i64>, <2 x i64>* %__b
-  %1 = bitcast <2 x i64> %load to <4 x i32>
-  %2 = icmp eq <4 x i32> %0, %1
-  %3 = bitcast i8 %__u to <8 x i1>
-  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
-  %4 = and <4 x i1> %2, %extract.i
-  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
-  %6 = bitcast <8 x i1> %5 to i8
-  ret i8 %6
-}
-
-
-define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
-; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
-; VLX:       # %bb.0: # %entry
-; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
-; VLX-NEXT:    kmovd %k0, %eax
-; VLX-NEXT:    # kill: def %al killed %al killed %eax
-; VLX-NEXT:    retq
-;
-; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
-; NoVLX:       # %bb.0: # %entry
-; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
-; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
-; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k0
-; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
-; NoVLX-NEXT:    andl $1, %eax
-; NoVLX-NEXT:    kmovw %eax, %k1
-; NoVLX-NEXT:    kxorw %k0, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
-; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
-; NoVLX-NEXT:    korw %k1, %k2, %k1
-; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
-; NoVLX-NEXT:    kxorw %k0, %k2, %k0
-; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
-; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
-; NoVLX-NEXT:    kxorw %k1, %k0, %k0
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
-; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
-; NoVLX-NEXT:    retq
-entry:
-  %0 = bitcast <2 x i64> %__a to <4 x i32>
+  %1 = bitcast <2 x i64> %__b to <4 x i32>
+  %2 = icmp eq <4 x i32> %0, %1
+  %3 = bitcast i8 %__u to <8 x i1>
+  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %4 = and <4 x i1> %2, %extract.i
+  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %6 = bitcast <8 x i1> %5 to i8
+  ret i8 %6
+}
+
+define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
+; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
+; VLX:       # %bb.0: # %entry
+; VLX-NEXT:    kmovd %edi, %k1
+; VLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    # kill: def %al killed %al killed %eax
+; VLX-NEXT:    retq
+;
+; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
+; NoVLX:       # %bb.0: # %entry
+; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
+; NoVLX-NEXT:    kmovw %edi, %k0
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
+; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
+; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k0
+; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
+; NoVLX-NEXT:    andl $1, %eax
+; NoVLX-NEXT:    kmovw %eax, %k1
+; NoVLX-NEXT:    kxorw %k0, %k0, %k2
+; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
+; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
+; NoVLX-NEXT:    korw %k1, %k2, %k1
+; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
+; NoVLX-NEXT:    kxorw %k0, %k2, %k0
+; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
+; NoVLX-NEXT:    kxorw %k1, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
+; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kmovw %k0, %eax
+; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
+; NoVLX-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__a to <4 x i32>
+  %load = load <2 x i64>, <2 x i64>* %__b
+  %1 = bitcast <2 x i64> %load to <4 x i32>
+  %2 = icmp eq <4 x i32> %0, %1
+  %3 = bitcast i8 %__u to <8 x i1>
+  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %4 = and <4 x i1> %2, %extract.i
+  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %6 = bitcast <8 x i1> %5 to i8
+  ret i8 %6
+}
+
+
+define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
+; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
+; VLX:       # %bb.0: # %entry
+; VLX-NEXT:    vpcmpeqd (%rdi){1to4}, %xmm0, %k0
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    # kill: def %al killed %al killed %eax
+; VLX-NEXT:    retq
+;
+; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
+; NoVLX:       # %bb.0: # %entry
+; NoVLX-NEXT:    vpbroadcastd (%rdi), %xmm1
+; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
+; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k0
+; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
+; NoVLX-NEXT:    andl $1, %eax
+; NoVLX-NEXT:    kmovw %eax, %k1
+; NoVLX-NEXT:    kxorw %k0, %k0, %k2
+; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
+; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
+; NoVLX-NEXT:    korw %k1, %k2, %k1
+; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
+; NoVLX-NEXT:    kxorw %k0, %k2, %k0
+; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
+; NoVLX-NEXT:    kxorw %k1, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
+; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kmovw %k0, %eax
+; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
+; NoVLX-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__a to <4 x i32>
   %load = load i32, i32* %__b
   %vec = insertelement <4 x i32> undef, i32 %load, i32 0
   %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
@@ -2673,15 +2673,15 @@ define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -2847,15 +2847,15 @@ define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -2915,15 +2915,15 @@ define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -3040,15 +3040,15 @@ define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -3200,15 +3200,15 @@ define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -3258,15 +3258,15 @@ define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -3363,15 +3363,15 @@ define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -3519,15 +3519,15 @@ define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -3583,15 +3583,15 @@ define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpeqd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -3700,15 +3700,15 @@ define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -3961,32 +3961,32 @@ define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4025,32 +4025,32 @@ define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>*
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4092,32 +4092,32 @@ define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4160,32 +4160,32 @@ define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4228,32 +4228,32 @@ define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4296,32 +4296,32 @@ define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4365,34 +4365,34 @@ define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4434,34 +4434,34 @@ define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>*
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4506,34 +4506,34 @@ define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4579,34 +4579,34 @@ define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4652,34 +4652,34 @@ define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -4725,34 +4725,34 @@ define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -6809,15 +6809,15 @@ define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -6880,15 +6880,15 @@ define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -7011,15 +7011,15 @@ define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -7194,15 +7194,15 @@ define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -7265,15 +7265,15 @@ define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -7396,15 +7396,15 @@ define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -7563,15 +7563,15 @@ define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -7623,15 +7623,15 @@ define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -7732,15 +7732,15 @@ define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -7894,15 +7894,15 @@ define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -7960,15 +7960,15 @@ define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -8081,15 +8081,15 @@ define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -8316,32 +8316,32 @@ define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8378,32 +8378,32 @@ define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>*
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8443,32 +8443,32 @@ define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8509,32 +8509,32 @@ define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8575,32 +8575,32 @@ define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8641,32 +8641,32 @@ define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8708,34 +8708,34 @@ define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8775,34 +8775,34 @@ define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>*
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpeqq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8845,34 +8845,34 @@ define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8916,34 +8916,34 @@ define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -8987,34 +8987,34 @@ define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpeqq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -9058,34 +9058,34 @@ define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -9824,32 +9824,32 @@ define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -9888,32 +9888,32 @@ define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -9955,32 +9955,32 @@ define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -10023,32 +10023,32 @@ define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -10091,34 +10091,34 @@ define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -10160,34 +10160,34 @@ define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -10232,34 +10232,34 @@ define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -10305,34 +10305,34 @@ define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -11590,15 +11590,15 @@ define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i6
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -11658,15 +11658,15 @@ define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -11783,15 +11783,15 @@ define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -11957,15 +11957,15 @@ define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -12025,15 +12025,15 @@ define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -12150,15 +12150,15 @@ define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -12310,15 +12310,15 @@ define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -12368,15 +12368,15 @@ define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -12473,15 +12473,15 @@ define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -12629,15 +12629,15 @@ define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -12693,15 +12693,15 @@ define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd (%rsi), %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -12810,15 +12810,15 @@ define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -13071,32 +13071,32 @@ define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13135,32 +13135,32 @@ define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13202,32 +13202,32 @@ define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13270,32 +13270,32 @@ define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13338,32 +13338,32 @@ define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13406,32 +13406,32 @@ define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13475,34 +13475,34 @@ define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13544,34 +13544,34 @@ define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13616,34 +13616,34 @@ define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13689,34 +13689,34 @@ define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13762,34 +13762,34 @@ define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -13835,34 +13835,34 @@ define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -15919,15 +15919,15 @@ define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i6
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -15990,15 +15990,15 @@ define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -16121,15 +16121,15 @@ define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -16304,15 +16304,15 @@ define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -16375,15 +16375,15 @@ define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -16506,15 +16506,15 @@ define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -16673,15 +16673,15 @@ define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -16733,15 +16733,15 @@ define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -16842,15 +16842,15 @@ define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -17004,15 +17004,15 @@ define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -17070,15 +17070,15 @@ define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -17191,15 +17191,15 @@ define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -17426,32 +17426,32 @@ define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17488,32 +17488,32 @@ define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17553,32 +17553,32 @@ define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17619,32 +17619,32 @@ define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17685,32 +17685,32 @@ define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17751,32 +17751,32 @@ define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17818,34 +17818,34 @@ define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17885,34 +17885,34 @@ define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpgtq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -17955,34 +17955,34 @@ define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -18026,34 +18026,34 @@ define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -18097,34 +18097,34 @@ define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpgtq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -18168,34 +18168,34 @@ define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -18976,32 +18976,32 @@ define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19043,32 +19043,32 @@ define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19112,32 +19112,32 @@ define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19183,32 +19183,32 @@ define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19253,34 +19253,34 @@ define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19325,34 +19325,34 @@ define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19399,34 +19399,34 @@ define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -19475,34 +19475,34 @@ define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -20801,84 +20801,84 @@ define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i6
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
-; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k0
-; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
-; NoVLX-NEXT:    andl $1, %eax
-; NoVLX-NEXT:    kmovw %eax, %k1
-; NoVLX-NEXT:    kxorw %k0, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
-; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
-; NoVLX-NEXT:    korw %k1, %k2, %k1
-; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
-; NoVLX-NEXT:    kxorw %k0, %k2, %k0
-; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
-; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
-; NoVLX-NEXT:    kxorw %k1, %k0, %k0
 ; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
-; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
-; NoVLX-NEXT:    kmovw %eax, %k2
-; NoVLX-NEXT:    kxorw %k2, %k1, %k1
-; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
-; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
-; NoVLX-NEXT:    kxorw %k0, %k1, %k0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
-; NoVLX-NEXT:    retq
-entry:
-  %0 = bitcast <2 x i64> %__a to <4 x i32>
-  %1 = bitcast <2 x i64> %__b to <4 x i32>
-  %2 = icmp sge <4 x i32> %0, %1
-  %3 = bitcast i8 %__u to <8 x i1>
-  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
-  %4 = and <4 x i1> %2, %extract.i
-  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
-  %6 = bitcast <8 x i1> %5 to i8
-  ret i8 %6
-}
-
-define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
-; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
-; VLX:       # %bb.0: # %entry
-; VLX-NEXT:    kmovd %edi, %k1
-; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
-; VLX-NEXT:    kmovd %k0, %eax
-; VLX-NEXT:    # kill: def %al killed %al killed %eax
-; VLX-NEXT:    retq
-;
-; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
-; NoVLX:       # %bb.0: # %entry
-; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
-; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
-; NoVLX-NEXT:    kmovw %edi, %k0
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
+; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
+; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k0
+; NoVLX-NEXT:    vpextrb $0, %xmm0, %eax
+; NoVLX-NEXT:    andl $1, %eax
+; NoVLX-NEXT:    kmovw %eax, %k1
+; NoVLX-NEXT:    kxorw %k0, %k0, %k2
+; NoVLX-NEXT:    kshiftrw $1, %k2, %k2
+; NoVLX-NEXT:    kshiftlw $1, %k2, %k2
+; NoVLX-NEXT:    korw %k1, %k2, %k1
+; NoVLX-NEXT:    kshiftrw $1, %k1, %k2
+; NoVLX-NEXT:    kxorw %k0, %k2, %k0
+; NoVLX-NEXT:    kshiftlw $15, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $14, %k0, %k0
+; NoVLX-NEXT:    kxorw %k1, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    vpextrb $8, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $13, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
+; NoVLX-NEXT:    vpextrb $12, %xmm0, %eax
+; NoVLX-NEXT:    kmovw %eax, %k2
+; NoVLX-NEXT:    kxorw %k2, %k1, %k1
+; NoVLX-NEXT:    kshiftlw $15, %k1, %k1
+; NoVLX-NEXT:    kshiftrw $12, %k1, %k1
+; NoVLX-NEXT:    kxorw %k0, %k1, %k0
+; NoVLX-NEXT:    kmovw %k0, %eax
+; NoVLX-NEXT:    # kill: def %al killed %al killed %eax
+; NoVLX-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__a to <4 x i32>
+  %1 = bitcast <2 x i64> %__b to <4 x i32>
+  %2 = icmp sge <4 x i32> %0, %1
+  %3 = bitcast i8 %__u to <8 x i1>
+  %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %4 = and <4 x i1> %2, %extract.i
+  %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %6 = bitcast <8 x i1> %5 to i8
+  ret i8 %6
+}
+
+define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
+; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
+; VLX:       # %bb.0: # %entry
+; VLX-NEXT:    kmovd %edi, %k1
+; VLX-NEXT:    vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
+; VLX-NEXT:    kmovd %k0, %eax
+; VLX-NEXT:    # kill: def %al killed %al killed %eax
+; VLX-NEXT:    retq
+;
+; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
+; NoVLX:       # %bb.0: # %entry
+; NoVLX-NEXT:    vmovdqa (%rsi), %xmm1
+; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
+; NoVLX-NEXT:    kmovw %edi, %k0
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -20997,15 +20997,15 @@ define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -21176,15 +21176,15 @@ define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -21245,15 +21245,15 @@ define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -21372,15 +21372,15 @@ define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -21537,15 +21537,15 @@ define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -21596,15 +21596,15 @@ define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -21703,15 +21703,15 @@ define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -21864,15 +21864,15 @@ define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -21929,15 +21929,15 @@ define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -22048,15 +22048,15 @@ define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpandn %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -22309,32 +22309,32 @@ define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22373,32 +22373,32 @@ define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22440,32 +22440,32 @@ define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22508,32 +22508,32 @@ define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22576,32 +22576,32 @@ define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22644,32 +22644,32 @@ define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22713,34 +22713,34 @@ define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22782,34 +22782,34 @@ define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22854,34 +22854,34 @@ define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -22927,34 +22927,34 @@ define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -23000,34 +23000,34 @@ define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -23073,34 +23073,34 @@ define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpled %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -25204,15 +25204,15 @@ define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i6
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25278,15 +25278,15 @@ define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25413,15 +25413,15 @@ define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25603,15 +25603,15 @@ define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25677,15 +25677,15 @@ define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25812,15 +25812,15 @@ define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -25986,15 +25986,15 @@ define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -26049,15 +26049,15 @@ define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -26162,15 +26162,15 @@ define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -26331,15 +26331,15 @@ define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -26400,15 +26400,15 @@ define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -26525,15 +26525,15 @@ define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -26760,32 +26760,32 @@ define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -26822,32 +26822,32 @@ define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -26887,32 +26887,32 @@ define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -26953,32 +26953,32 @@ define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27019,32 +27019,32 @@ define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27085,32 +27085,32 @@ define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27152,34 +27152,34 @@ define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27219,34 +27219,34 @@ define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpnltq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27289,34 +27289,34 @@ define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpleq %zmm0, %zmm1, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27360,34 +27360,34 @@ define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27431,34 +27431,34 @@ define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpnltq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -27502,34 +27502,34 @@ define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28319,32 +28319,32 @@ define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28386,32 +28386,32 @@ define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28456,32 +28456,32 @@ define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28527,32 +28527,32 @@ define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28598,34 +28598,34 @@ define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28670,34 +28670,34 @@ define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>
 ; NoVLX-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28745,34 +28745,34 @@ define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -28821,34 +28821,34 @@ define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vptestmq %zmm0, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -30159,15 +30159,15 @@ define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i6
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30230,15 +30230,15 @@ define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30361,15 +30361,15 @@ define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30544,15 +30544,15 @@ define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30615,15 +30615,15 @@ define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30746,15 +30746,15 @@ define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -30915,15 +30915,15 @@ define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -30976,15 +30976,15 @@ define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -31087,15 +31087,15 @@ define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -31252,15 +31252,15 @@ define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -31319,15 +31319,15 @@ define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -31442,15 +31442,15 @@ define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpcmpgtd %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -31703,32 +31703,32 @@ define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -31767,32 +31767,32 @@ define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -31834,32 +31834,32 @@ define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -31902,32 +31902,32 @@ define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -31970,32 +31970,32 @@ define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32038,32 +32038,32 @@ define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32107,34 +32107,34 @@ define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32176,34 +32176,34 @@ define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovdqa (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32248,34 +32248,34 @@ define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32321,34 +32321,34 @@ define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovdqa (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32394,34 +32394,34 @@ define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %_
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vpbroadcastd (%rdi), %ymm1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -32467,34 +32467,34 @@ define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpbroadcastd (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -34650,15 +34650,15 @@ define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i6
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -34724,15 +34724,15 @@ define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -34861,15 +34861,15 @@ define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -35053,15 +35053,15 @@ define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -35127,15 +35127,15 @@ define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -35264,15 +35264,15 @@ define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpextrb $4, %xmm0, %eax
@@ -35440,15 +35440,15 @@ define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -35503,15 +35503,15 @@ define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -35618,15 +35618,15 @@ define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -35789,15 +35789,15 @@ define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -35858,15 +35858,15 @@ define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -35985,15 +35985,15 @@ define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vpmovqd %zmm0, %ymm0
 ; NoVLX-NEXT:    kmovw %edi, %k0
 ; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k2
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k3
-; NoVLX-NEXT:    kmovw %k3, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    vmovd %ecx, %xmm1
-; NoVLX-NEXT:    vpinsrb $4, %eax, %xmm1, %xmm1
-; NoVLX-NEXT:    kmovw %k2, %eax
-; NoVLX-NEXT:    vpinsrb $8, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    kmovw %k1, %eax
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kmovw %k0, %esi
+; NoVLX-NEXT:    vmovd %esi, %xmm1
+; NoVLX-NEXT:    vpinsrb $4, %edx, %xmm1, %xmm1
+; NoVLX-NEXT:    vpinsrb $8, %ecx, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpinsrb $12, %eax, %xmm1, %xmm1
 ; NoVLX-NEXT:    vpand %xmm0, %xmm1, %xmm0
 ; NoVLX-NEXT:    kxorw %k0, %k0, %k0
@@ -36220,32 +36220,32 @@ define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36282,32 +36282,32 @@ define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36347,32 +36347,32 @@ define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36413,32 +36413,32 @@ define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36479,32 +36479,32 @@ define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36545,32 +36545,32 @@ define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36612,34 +36612,34 @@ define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36679,34 +36679,34 @@ define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpltuq (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36749,34 +36749,34 @@ define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36820,34 +36820,34 @@ define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36891,34 +36891,34 @@ define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %_
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vpcmpltuq (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -36962,34 +36962,34 @@ define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38525,32 +38525,32 @@ define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38589,32 +38589,32 @@ define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovaps (%rdi), %ymm1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38654,32 +38654,32 @@ define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, float*
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vbroadcastss (%rdi), %ymm1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38722,32 +38722,32 @@ define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38790,32 +38790,32 @@ define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovaps (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38859,32 +38859,32 @@ define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vbroadcastss (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38929,34 +38929,34 @@ define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__
 ; NoVLX-NEXT:    # kill: def %ymm1 killed %ymm1 def %zmm1
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -38998,34 +38998,34 @@ define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vmovaps (%rdi), %ymm1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -39068,34 +39068,34 @@ define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, float*
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    vbroadcastss (%rdi), %ymm1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -39141,34 +39141,34 @@ define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask(i8 zeroext %__u, <4 x
 ; NoVLX-NEXT:    # kill: def %ymm0 killed %ymm0 def %zmm0
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -39214,34 +39214,34 @@ define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    vmovaps (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -39288,34 +39288,34 @@ define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    vbroadcastss (%rsi), %ymm1
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqps %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorps %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -42910,32 +42910,32 @@ define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -42972,32 +42972,32 @@ define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43035,32 +43035,32 @@ define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, double*
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43101,32 +43101,32 @@ define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43167,32 +43167,32 @@ define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43234,32 +43234,32 @@ define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $32, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43352,34 +43352,34 @@ define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43419,34 +43419,34 @@ define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vcmpeqpd (%rdi), %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43487,34 +43487,34 @@ define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, double*
 ; NoVLX-NEXT:    andq $-32, %rsp
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    vcmpeqpd (%rdi){1to8}, %zmm0, %k0
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43558,34 +43558,34 @@ define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask(i8 zeroext %__u, <8 x
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43629,34 +43629,34 @@ define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -43701,34 +43701,34 @@ define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u,
 ; NoVLX-NEXT:    subq $64, %rsp
 ; NoVLX-NEXT:    kmovw %edi, %k1
 ; NoVLX-NEXT:    vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
-; NoVLX-NEXT:    kxorw %k0, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kmovw %k1, {{[0-9]+}}(%rsp)
-; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $7, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r8d
-; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %r9d
-; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
-; NoVLX-NEXT:    kmovw %k1, %edx
+; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %r10d
 ; NoVLX-NEXT:    kshiftrw $4, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %esi
-; NoVLX-NEXT:    kshiftrw $5, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $3, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %edi
-; NoVLX-NEXT:    kshiftrw $6, %k0, %k1
+; NoVLX-NEXT:    kshiftrw $2, %k0, %k1
 ; NoVLX-NEXT:    kmovw %k1, %eax
-; NoVLX-NEXT:    kmovw %k0, %ecx
-; NoVLX-NEXT:    kshiftrw $7, %k0, %k0
+; NoVLX-NEXT:    kshiftrw $1, %k0, %k1
+; NoVLX-NEXT:    kmovw %k1, %ecx
+; NoVLX-NEXT:    kmovw %k0, %edx
+; NoVLX-NEXT:    kxorw %k0, %k0, %k0
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
+; NoVLX-NEXT:    kmovw %k0, {{[0-9]+}}(%rsp)
 ; NoVLX-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $0, %ecx, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $1, %r8d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $2, %r9d, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $3, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $0, %edx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $1, %ecx, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $2, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $3, %edi, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpinsrb $4, %esi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $5, %edi, %xmm0, %xmm0
-; NoVLX-NEXT:    vpinsrb $6, %eax, %xmm0, %xmm0
-; NoVLX-NEXT:    kmovw %k0, %eax
-; NoVLX-NEXT:    vpinsrb $7, %eax, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $5, %r10d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $6, %r9d, %xmm0, %xmm0
+; NoVLX-NEXT:    vpinsrb $7, %r8d, %xmm0, %xmm0
 ; NoVLX-NEXT:    vpmovsxbd %xmm0, %zmm0
 ; NoVLX-NEXT:    vpslld $31, %zmm0, %zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0