diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18661,6 +18661,13 @@
 
   // Allow targets to opt-out.
   EVT VT = Extract->getValueType(0);
+
+  // Only handle byte sized scalars otherwise the offset is incorrect.
+  // FIXME: We might be able to do better if the VT is byte sized and the index
+  // is aligned.
+  if (!VT.getScalarType().isByteSized())
+    return SDValue();
+
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   if (!TLI.shouldReduceLoadWidth(Ld, Ld->getExtensionType(), VT))
     return SDValue();
diff --git a/llvm/test/CodeGen/X86/avx512-extract-subvector-load-store.ll b/llvm/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
--- a/llvm/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
+++ b/llvm/test/CodeGen/X86/avx512-extract-subvector-load-store.ll
@@ -5,7 +5,8 @@
 define void @load_v8i1_broadcast_4_v2i1(<8 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v8i1_broadcast_4_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 4(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastq %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -15,7 +16,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_4_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $4, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm2, %xmm2
@@ -32,7 +34,8 @@
 define void @load_v8i1_broadcast_7_v2i1(<8 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v8i1_broadcast_7_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 6(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $6, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -42,7 +45,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_7_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 6(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $6, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
@@ -59,7 +63,8 @@
 define void @load_v16i1_broadcast_8_v2i1(<16 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v16i1_broadcast_8_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 8(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastq %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -69,7 +74,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_8_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 8(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $8, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm2, %xmm2
@@ -86,7 +92,8 @@
 define void @load_v16i1_broadcast_8_v4i1(<16 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v16i1_broadcast_8_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 8(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -96,7 +103,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_8_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 8(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $8, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %xmm2
@@ -113,7 +121,8 @@
 define void @load_v16i1_broadcast_15_v2i1(<16 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v16i1_broadcast_15_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 14(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $14, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -123,7 +132,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_15_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 14(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $14, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
@@ -140,7 +150,8 @@
 define void @load_v16i1_broadcast_15_v4i1(<16 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v16i1_broadcast_15_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 12(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $12, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -150,7 +161,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_15_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 12(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $12, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
@@ -167,7 +179,8 @@
 define void @load_v32i1_broadcast_16_v2i1(<32 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastq %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -177,7 +190,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm2, %xmm2
@@ -194,7 +208,8 @@
 define void @load_v32i1_broadcast_16_v4i1(<32 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -204,7 +219,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %xmm2
@@ -221,7 +237,8 @@
 define void @load_v32i1_broadcast_16_v8i1(<32 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
@@ -232,7 +249,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %ymm2
@@ -250,7 +268,8 @@
 define void @load_v32i1_broadcast_31_v2i1(<32 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 30(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $30, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -260,7 +279,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 30(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $30, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
@@ -277,7 +297,8 @@
 define void @load_v32i1_broadcast_31_v4i1(<32 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 28(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $28, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -287,7 +308,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 28(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $28, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
@@ -304,7 +326,8 @@
 define void @load_v32i1_broadcast_31_v8i1(<32 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 24(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $24, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm2, %ymm3, %ymm2
@@ -316,7 +339,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 24(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $24, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
@@ -335,7 +359,8 @@
 define void @load_v64i1_broadcast_32_v2i1(<64 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastq %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -345,7 +370,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm2, %xmm2
@@ -362,7 +388,8 @@
 define void @load_v64i1_broadcast_32_v4i1(<64 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -372,7 +399,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %xmm2
@@ -389,7 +417,8 @@
 define void @load_v64i1_broadcast_32_v8i1(<64 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
@@ -400,7 +429,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %ymm2
@@ -418,7 +448,8 @@
 define void @load_v64i1_broadcast_32_v16i1(<64 x i1>* %a0,<16 x float> %a1,<16 x float> %a2,<16 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v16i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovw 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %zmm2
 ; AVX512-NEXT:    vpmovd2m %zmm2, %k1
@@ -429,7 +460,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v16i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %zmm2
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm2, %zmm2, %k1
@@ -446,7 +478,8 @@
 define void @load_v64i1_broadcast_63_v2i1(<64 x i1>* %a0,<2 x double> %a1,<2 x double> %a2,<2 x double>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v2i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 62(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $62, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm2, %k1
@@ -456,7 +489,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v2i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 62(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $62, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
@@ -473,7 +507,8 @@
 define void @load_v64i1_broadcast_63_v4i1(<64 x i1>* %a0,<4 x float> %a1,<4 x float> %a2,<4 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v4i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 60(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $60, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm2
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm2, %k1
@@ -483,7 +518,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v4i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 60(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $60, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm2, %xmm2, %xmm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm2, %xmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
@@ -500,7 +536,8 @@
 define void @load_v64i1_broadcast_63_v8i1(<64 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 56(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $56, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm2, %ymm3, %ymm2
@@ -512,7 +549,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 56(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $56, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
@@ -531,7 +569,8 @@
 define void @load_v64i1_broadcast_63_v16i1(<64 x i1>* %a0,<16 x float> %a1,<16 x float> %a2,<16 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v16i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovw 48(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $48, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512-NEXT:    vpermd %zmm2, %zmm3, %zmm2
@@ -543,7 +582,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v16i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 48(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $48, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512NOTDQ-NEXT:    vpermd %zmm2, %zmm3, %zmm2
@@ -561,13 +601,16 @@
 define void @load_v2i1_broadcast_1_v1i1_store(<2 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v2i1_broadcast_1_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 1(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $1, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v2i1_broadcast_1_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 1(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $1, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <2 x i1>, <2 x i1>* %a0
@@ -578,13 +621,26 @@
 define void @load_v3i1_broadcast_1_v1i1_store(<3 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v3i1_broadcast_1_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 1(%rdi), %k0
+; AVX512-NEXT:    movzbl (%rdi), %eax
+; AVX512-NEXT:    xorl %ecx, %ecx
+; AVX512-NEXT:    btl $1, %eax
+; AVX512-NEXT:    movl $255, %eax
+; AVX512-NEXT:    cmovael %ecx, %eax
+; AVX512-NEXT:    kmovd %eax, %k0
+; AVX512-NEXT:    kshiftrb $1, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v3i1_broadcast_1_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 1(%rdi), %al
+; AVX512NOTDQ-NEXT:    movzbl (%rdi), %eax
+; AVX512NOTDQ-NEXT:    xorl %ecx, %ecx
+; AVX512NOTDQ-NEXT:    btl $1, %eax
+; AVX512NOTDQ-NEXT:    movl $255, %eax
+; AVX512NOTDQ-NEXT:    cmovael %ecx, %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $1, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <3 x i1>, <3 x i1>* %a0
@@ -595,13 +651,26 @@
 define void @load_v3i1_broadcast_2_v1i1_store(<3 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v3i1_broadcast_2_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 2(%rdi), %k0
+; AVX512-NEXT:    movzbl (%rdi), %eax
+; AVX512-NEXT:    xorl %ecx, %ecx
+; AVX512-NEXT:    btl $2, %eax
+; AVX512-NEXT:    movl $255, %eax
+; AVX512-NEXT:    cmovael %ecx, %eax
+; AVX512-NEXT:    kmovd %eax, %k0
+; AVX512-NEXT:    kshiftrb $2, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v3i1_broadcast_2_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 2(%rdi), %al
+; AVX512NOTDQ-NEXT:    movzbl (%rdi), %eax
+; AVX512NOTDQ-NEXT:    xorl %ecx, %ecx
+; AVX512NOTDQ-NEXT:    btl $2, %eax
+; AVX512NOTDQ-NEXT:    movl $255, %eax
+; AVX512NOTDQ-NEXT:    cmovael %ecx, %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $2, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <3 x i1>, <3 x i1>* %a0
@@ -612,13 +681,16 @@
 define void @load_v4i1_broadcast_2_v1i1_store(<4 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v4i1_broadcast_2_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 2(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $2, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v4i1_broadcast_2_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 2(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $2, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <4 x i1>, <4 x i1>* %a0
@@ -629,13 +701,16 @@
 define void @load_v4i1_broadcast_3_v1i1_store(<4 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v4i1_broadcast_3_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 3(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $3, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v4i1_broadcast_3_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 3(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $3, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <4 x i1>, <4 x i1>* %a0
@@ -646,13 +721,16 @@
 define void @load_v8i1_broadcast_4_v1i1_store(<8 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_4_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 4(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_4_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 4(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $4, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <8 x i1>, <8 x i1>* %a0
@@ -663,7 +741,8 @@
 define void @load_v8i1_broadcast_4_v2i1_store(<8 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_4_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 4(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $4, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastq %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -672,7 +751,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_4_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $4, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm0, %xmm0
@@ -688,13 +768,16 @@
 define void @load_v8i1_broadcast_7_v1i1_store(<8 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_7_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 7(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $7, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_7_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 7(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $7, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <8 x i1>, <8 x i1>* %a0
@@ -705,7 +788,8 @@
 define void @load_v8i1_broadcast_7_v2i1_store(<8 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v8i1_broadcast_7_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 6(%rdi), %k0
+; AVX512-NEXT:    kmovb (%rdi), %k0
+; AVX512-NEXT:    kshiftrb $6, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -714,7 +798,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v8i1_broadcast_7_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 6(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $6, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
@@ -730,13 +815,16 @@
 define void @load_v16i1_broadcast_8_v1i1_store(<16 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_8_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 8(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_8_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 8(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $8, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <16 x i1>, <16 x i1>* %a0
@@ -747,7 +835,8 @@
 define void @load_v16i1_broadcast_8_v2i1_store(<16 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_8_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 8(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastq %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -756,7 +845,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_8_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 8(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $8, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm0, %xmm0
@@ -772,7 +862,8 @@
 define void @load_v16i1_broadcast_8_v4i1_store(<16 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_8_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 8(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -781,7 +872,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_8_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 8(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $8, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %xmm0
@@ -797,13 +889,16 @@
 define void @load_v16i1_broadcast_15_v1i1_store(<16 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_15_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 15(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $15, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_15_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 15(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $15, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <16 x i1>, <16 x i1>* %a0
@@ -814,7 +909,8 @@
 define void @load_v16i1_broadcast_15_v2i1_store(<16 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_15_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 14(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $14, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -823,7 +919,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_15_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 14(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $14, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
@@ -839,7 +936,8 @@
 define void @load_v16i1_broadcast_15_v4i1_store(<16 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v16i1_broadcast_15_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 12(%rdi), %k0
+; AVX512-NEXT:    kmovw (%rdi), %k0
+; AVX512-NEXT:    kshiftrw $12, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -848,7 +946,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v16i1_broadcast_15_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 12(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovw (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrw $12, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
@@ -864,13 +963,16 @@
 define void @load_v32i1_broadcast_16_v1i1_store(<32 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 16(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <32 x i1>, <32 x i1>* %a0
@@ -881,7 +983,8 @@
 define void @load_v32i1_broadcast_16_v2i1_store(<32 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastq %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -890,7 +993,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm0, %xmm0
@@ -906,7 +1010,8 @@
 define void @load_v32i1_broadcast_16_v4i1_store(<32 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -915,7 +1020,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %xmm0
@@ -931,7 +1037,8 @@
 define void @load_v32i1_broadcast_16_v8i1_store(<32 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 16(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $16, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
@@ -941,7 +1048,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 16(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %ymm0
@@ -958,13 +1066,16 @@
 define void @load_v32i1_broadcast_31_v1i1_store(<32 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 31(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $31, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 31(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $31, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <32 x i1>, <32 x i1>* %a0
@@ -975,7 +1086,8 @@
 define void @load_v32i1_broadcast_31_v2i1_store(<32 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 30(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $30, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -984,7 +1096,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 30(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $30, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
@@ -1000,7 +1113,8 @@
 define void @load_v32i1_broadcast_31_v4i1_store(<32 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 28(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $28, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -1009,7 +1123,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 28(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $28, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
@@ -1025,7 +1140,8 @@
 define void @load_v32i1_broadcast_31_v8i1_store(<32 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 24(%rdi), %k0
+; AVX512-NEXT:    kmovd (%rdi), %k0
+; AVX512-NEXT:    kshiftrd $24, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm0, %ymm1, %ymm0
@@ -1036,7 +1152,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 24(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrd $24, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
@@ -1054,13 +1171,16 @@
 define void @load_v64i1_broadcast_32_v1i1_store(<64 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 32(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
@@ -1071,7 +1191,8 @@
 define void @load_v64i1_broadcast_32_v2i1_store(<64 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastq %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -1080,7 +1201,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastq %xmm0, %xmm0
@@ -1096,7 +1218,8 @@
 define void @load_v64i1_broadcast_32_v4i1_store(<64 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -1105,7 +1228,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %xmm0
@@ -1121,7 +1245,8 @@
 define void @load_v64i1_broadcast_32_v8i1_store(<64 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
@@ -1131,7 +1256,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %ymm0
@@ -1148,7 +1274,8 @@
 define void @load_v64i1_broadcast_32_v16i1_store(<64 x i1>* %a0,<16 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v16i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovw 32(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $32, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %zmm0
 ; AVX512-NEXT:    vpmovd2m %zmm0, %k0
@@ -1158,7 +1285,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v16i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 32(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %zmm0
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm0, %zmm0, %k0
@@ -1173,13 +1301,16 @@
 define void @load_v64i1_broadcast_63_v1i1_store(<64 x i1>* %a0,<1 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v1i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 63(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $63, %k0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v1i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    movb 63(%rdi), %al
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $63, %k0, %k0
+; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
@@ -1190,7 +1321,8 @@
 define void @load_v64i1_broadcast_63_v2i1_store(<64 x i1>* %a0,<2 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v2i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 62(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $62, %k0, %k0
 ; AVX512-NEXT:    vpmovm2q %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX512-NEXT:    vpmovq2m %xmm0, %k0
@@ -1199,7 +1331,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v2i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 62(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $62, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa64 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
@@ -1215,7 +1348,8 @@
 define void @load_v64i1_broadcast_63_v4i1_store(<64 x i1>* %a0,<4 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v4i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 60(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $60, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
 ; AVX512-NEXT:    vpmovd2m %xmm0, %k0
@@ -1224,7 +1358,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v4i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 60(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $60, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %xmm0, %xmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
@@ -1240,7 +1375,8 @@
 define void @load_v64i1_broadcast_63_v8i1_store(<64 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovb 56(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $56, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm0, %ymm1, %ymm0
@@ -1251,7 +1387,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 56(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $56, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
@@ -1269,7 +1406,8 @@
 define void @load_v64i1_broadcast_63_v16i1_store(<64 x i1>* %a0,<16 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v16i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovw 48(%rdi), %k0
+; AVX512-NEXT:    kmovq (%rdi), %k0
+; AVX512-NEXT:    kshiftrq $48, %k0, %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512-NEXT:    vpermd %zmm0, %zmm1, %zmm0
@@ -1280,7 +1418,8 @@
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v16i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovw 48(%rdi), %k1
+; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
+; AVX512NOTDQ-NEXT:    kshiftrq $48, %k0, %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512NOTDQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
diff --git a/llvm/test/CodeGen/X86/avx512-mask-op.ll b/llvm/test/CodeGen/X86/avx512-mask-op.ll
--- a/llvm/test/CodeGen/X86/avx512-mask-op.ll
+++ b/llvm/test/CodeGen/X86/avx512-mask-op.ll
@@ -278,7 +278,10 @@
 ;
 ; X86-LABEL: shuf_test1:
 ; X86:       ## %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k0
+; X86-NEXT:    kshiftrw $8, %k0, %k0
+; X86-NEXT:    kmovd %k0, %eax
+; X86-NEXT:    ## kill: def $al killed $al killed $eax
 ; X86-NEXT:    retl
    %v1 = bitcast i16 %v to <16 x i1>
    %mask = shufflevector <16 x i1> %v1, <16 x i1> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
diff --git a/llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll b/llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll
--- a/llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll
+++ b/llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll
@@ -7,8 +7,8 @@
 define i32@test_int_x86_avx512_kunpck_wd(i32 %x0, i32 %x1) {
 ; X86-LABEL: test_int_x86_avx512_kunpck_wd:
 ; X86:       # %bb.0:
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k0 # encoding: [0xc5,0xf8,0x90,0x44,0x24,0x04]
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 # encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x08]
+; X86-NEXT:    kmovd {{[0-9]+}}(%esp), %k0 # encoding: [0xc4,0xe1,0xf9,0x90,0x44,0x24,0x04]
+; X86-NEXT:    kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
 ; X86-NEXT:    kunpckwd %k1, %k0, %k0 # encoding: [0xc5,0xfc,0x4b,0xc1]
 ; X86-NEXT:    kmovd %k0, %eax # encoding: [0xc5,0xfb,0x93,0xc0]
 ; X86-NEXT:    retl # encoding: [0xc3]
@@ -29,8 +29,10 @@
 define i64@test_int_x86_avx512_kunpck_qd(i64 %x0, i64 %x1) {
 ; X86-LABEL: test_int_x86_avx512_kunpck_qd:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx # encoding: [0x8b,0x54,0x24,0x04]
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x0c]
+; X86-NEXT:    kmovq {{[0-9]+}}(%esp), %k0 # encoding: [0xc4,0xe1,0xf8,0x90,0x44,0x24,0x04]
+; X86-NEXT:    kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x0c]
+; X86-NEXT:    kmovd %k0, %edx # encoding: [0xc5,0xfb,0x93,0xd0]
+; X86-NEXT:    kmovd %k1, %eax # encoding: [0xc5,0xfb,0x93,0xc1]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_kunpck_qd: