[llvm] 2cf31ba - [X86] combineConcatVectorOps - add X86ISD::VPERMV handling
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Mon Feb 6 07:42:14 PST 2023
Author: Simon Pilgrim
Date: 2023-02-06T15:41:59Z
New Revision: 2cf31ba5dccecd48d5f3c9d9d5c3808a144a7e6e
URL: https://github.com/llvm/llvm-project/commit/2cf31ba5dccecd48d5f3c9d9d5c3808a144a7e6e
DIFF: https://github.com/llvm/llvm-project/commit/2cf31ba5dccecd48d5f3c9d9d5c3808a144a7e6e.diff
LOG: [X86] combineConcatVectorOps - add X86ISD::VPERMV handling
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 90240926d3ae..c266c1872e6e 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -55390,6 +55390,34 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
ConcatSubOperand(VT, Ops, 1));
}
break;
+ case X86ISD::VPERMV:
+ if (!IsSplat && NumOps == 2 &&
+ (VT.is512BitVector() && Subtarget.useAVX512Regs())) {
+ MVT OpVT = Op0.getSimpleValueType();
+ int NumSrcElts = OpVT.getVectorNumElements();
+ SmallVector<int, 64> ConcatMask;
+ for (unsigned i = 0; i != NumOps; ++i) {
+ SmallVector<int, 64> SubMask;
+ SmallVector<SDValue, 2> SubOps;
+ if (!getTargetShuffleMask(Ops[i].getNode(), OpVT, false, SubOps,
+ SubMask))
+ break;
+ for (int M : SubMask) {
+ if (0 <= M)
+ M += i * NumSrcElts;
+ ConcatMask.push_back(M);
+ }
+ }
+ if (ConcatMask.size() == (NumOps * NumSrcElts)) {
+ SDValue Src = concatSubVectors(Ops[0].getOperand(1),
+ Ops[1].getOperand(1), DAG, DL);
+ MVT IntMaskSVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
+ MVT IntMaskVT = MVT::getVectorVT(IntMaskSVT, NumOps * NumSrcElts);
+ SDValue Mask = getConstVector(ConcatMask, IntMaskVT, DAG, DL, true);
+ return DAG.getNode(X86ISD::VPERMV, DL, VT, Mask, Src);
+ }
+ }
+ break;
case X86ISD::VPERMV3:
if (!IsSplat && NumOps == 2 && VT.is512BitVector()) {
MVT OpVT = Op0.getSimpleValueType();
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
index 017c8b7ea44a..eb0ef5caaa0a 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll
@@ -1215,61 +1215,60 @@ define void @store_i16_stride3_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
; AVX512F-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
; AVX512F-NEXT: vinserti128 $1, %xmm6, %ymm3, %ymm3
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512F-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = <u,0,0,u,1,1,u,2>
-; AVX512F-NEXT: vpermd %ymm5, %ymm6, %ymm8
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm4 = <5,5,u,6,6,u,7,7,u,8,8,u,9,9,u,10>
+; AVX512F-NEXT: vpermd (%rdx), %zmm4, %zmm4
+; AVX512F-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512F-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512F-NEXT: vpshufb %ymm5, %ymm3, %ymm3
+; AVX512F-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
+; AVX512F-NEXT: vpshufb %ymm8, %ymm6, %ymm6
+; AVX512F-NEXT: vpor %ymm3, %ymm6, %ymm3
+; AVX512F-NEXT: vmovdqa 48(%rdi), %xmm6
+; AVX512F-NEXT: vmovdqa 48(%rsi), %xmm9
+; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512F-NEXT: vpshufb %xmm7, %xmm10, %xmm7
+; AVX512F-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm9[2],xmm6[3,4],xmm9[5],xmm6[6,7]
+; AVX512F-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6
+; AVX512F-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm6[4,5,6,7]
+; AVX512F-NEXT: vmovdqa (%rdx), %ymm6
+; AVX512F-NEXT: vmovdqa 32(%rdx), %ymm7
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm9 = <5,5,u,6,6,u,7,7>
-; AVX512F-NEXT: vpermd %ymm3, %ymm9, %ymm10
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512F-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm8
-; AVX512F-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512F-NEXT: vpshufb %ymm10, %ymm4, %ymm4
-; AVX512F-NEXT: vmovdqa 32(%rsi), %ymm11
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm12 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
-; AVX512F-NEXT: vpshufb %ymm12, %ymm11, %ymm11
-; AVX512F-NEXT: vpor %ymm4, %ymm11, %ymm4
-; AVX512F-NEXT: vmovdqa 48(%rdi), %xmm11
-; AVX512F-NEXT: vmovdqa 48(%rsi), %xmm13
-; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512F-NEXT: vpshufb %xmm7, %xmm14, %xmm7
-; AVX512F-NEXT: vprold $16, %xmm13, %xmm13
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm11 = xmm11[0,1],xmm13[2],xmm11[3,4],xmm13[5],xmm11[6,7]
-; AVX512F-NEXT: vinserti128 $1, %xmm7, %ymm11, %ymm7
-; AVX512F-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm7[4,5,6,7]
-; AVX512F-NEXT: vpermd %ymm5, %ymm9, %ymm7
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm9 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512F-NEXT: vpandn %ymm7, %ymm9, %ymm7
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512F-NEXT: vpshufb %ymm9, %ymm5, %ymm5
-; AVX512F-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
-; AVX512F-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
-; AVX512F-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512F-NEXT: vpshufb %ymm10, %ymm4, %ymm4
-; AVX512F-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512F-NEXT: vpshufb %ymm12, %ymm7, %ymm7
-; AVX512F-NEXT: vpor %ymm4, %ymm7, %ymm4
-; AVX512F-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512F-NEXT: vprold $16, %xmm0, %xmm7
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm10 = xmm1[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm7 = xmm10[0,1],xmm7[2],xmm10[3,4],xmm7[5],xmm10[6,7]
+; AVX512F-NEXT: vpermd %ymm7, %ymm9, %ymm9
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm10 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512F-NEXT: vpandn %ymm9, %ymm10, %ymm9
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512F-NEXT: vpshufb %ymm10, %ymm7, %ymm7
+; AVX512F-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512F-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm7
+; AVX512F-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512F-NEXT: vpshufb %ymm5, %ymm3, %ymm3
+; AVX512F-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512F-NEXT: vpshufb %ymm8, %ymm5, %ymm5
+; AVX512F-NEXT: vpor %ymm3, %ymm5, %ymm3
+; AVX512F-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
+; AVX512F-NEXT: vprold $16, %xmm0, %xmm5
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm1[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512F-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512F-NEXT: vinserti128 $1, %xmm7, %ymm0, %ymm0
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
-; AVX512F-NEXT: vpshufb %ymm9, %ymm3, %ymm1
-; AVX512F-NEXT: vpermd %ymm3, %ymm6, %ymm2
+; AVX512F-NEXT: vinserti128 $1, %xmm5, %ymm0, %ymm0
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
+; AVX512F-NEXT: vpshufb %ymm10, %ymm6, %ymm1
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm2 = <u,0,0,u,1,1,u,2>
+; AVX512F-NEXT: vpermd %ymm6, %ymm2, %ymm2
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX512F-NEXT: vpandn %ymm2, %ymm3, %ymm2
; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512F-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm1, (%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm5, 128(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm8, 64(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm7, 128(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm4, 64(%rcx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
@@ -2332,158 +2331,154 @@ define void @store_i16_stride3_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-LABEL: store_i16_stride3_vf64:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
-; AVX512F-NEXT: vpshufb %ymm6, %ymm0, %ymm0
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128,20,21,128,128]
+; AVX512F-NEXT: vpshufb %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vmovdqa %ymm1, %ymm7
; AVX512F-NEXT: vmovdqa (%rsi), %ymm1
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm2 = [10,11,0,1,128,128,12,13,2,3,128,128,14,15,4,5,128,128,16,17,28,29,128,128,18,19,18,19,128,128,20,21]
; AVX512F-NEXT: vpshufb %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT: vmovdqa %ymm2, %ymm10
; AVX512F-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm2
-; AVX512F-NEXT: vmovdqa (%rsi), %xmm4
-; AVX512F-NEXT: vmovdqa64 16(%rsi), %xmm23
-; AVX512F-NEXT: vmovdqa64 32(%rsi), %xmm25
-; AVX512F-NEXT: vprold $16, %xmm4, %xmm5
-; AVX512F-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512F-NEXT: vmovdqa64 16(%rdi), %xmm24
-; AVX512F-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
-; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
+; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm3
+; AVX512F-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512F-NEXT: vmovdqa64 16(%rsi), %xmm24
+; AVX512F-NEXT: vmovdqa 32(%rsi), %xmm6
+; AVX512F-NEXT: vprold $16, %xmm5, %xmm8
+; AVX512F-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512F-NEXT: vmovdqa64 16(%rdi), %xmm25
+; AVX512F-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
+; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
; AVX512F-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512F-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm4[0,1,2,3],zmm2[4,5,6,7]
-; AVX512F-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512F-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512F-NEXT: vmovdqa 64(%rdx), %ymm15
-; AVX512F-NEXT: vmovdqa 96(%rdx), %ymm7
-; AVX512F-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
-; AVX512F-NEXT: vpshufb %ymm3, %ymm2, %ymm4
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm20 = <u,0,0,u,1,1,u,2>
-; AVX512F-NEXT: vpermd %ymm2, %ymm20, %ymm13
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm17 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512F-NEXT: vpandnq %ymm13, %ymm17, %ymm13
-; AVX512F-NEXT: vinserti64x4 $1, %ymm4, %zmm13, %zmm4
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512F-NEXT: vpternlogq $248, %zmm18, %zmm12, %zmm4
-; AVX512F-NEXT: vmovdqa 96(%rsi), %xmm12
-; AVX512F-NEXT: vprold $16, %xmm12, %xmm13
-; AVX512F-NEXT: vmovdqa 96(%rdi), %xmm14
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm14[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
-; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512F-NEXT: vpshufb %xmm0, %xmm12, %xmm12
+; AVX512F-NEXT: vpshufb %xmm0, %xmm9, %xmm9
+; AVX512F-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm8[0,1,2,3],zmm3[4,5,6,7]
+; AVX512F-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512F-NEXT: vmovdqa 32(%rdx), %ymm8
+; AVX512F-NEXT: vmovdqa 64(%rdx), %ymm14
+; AVX512F-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,10,11,128,128,128,128,12,13,128,128,128,128,14,15,128,128,128,128,16,17,128,128,128,128,18,19,128,128,128,128]
+; AVX512F-NEXT: vpshufb %ymm9, %ymm3, %ymm11
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm19 = <u,0,0,u,1,1,u,2>
+; AVX512F-NEXT: vpermd %ymm3, %ymm19, %ymm3
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512F-NEXT: vpandnq %ymm3, %ymm16, %ymm3
+; AVX512F-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512F-NEXT: vpternlogq $248, %zmm17, %zmm10, %zmm3
+; AVX512F-NEXT: vmovdqa 96(%rsi), %xmm10
+; AVX512F-NEXT: vprold $16, %xmm10, %xmm11
+; AVX512F-NEXT: vmovdqa 96(%rdi), %xmm12
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm13 = xmm12[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm11 = xmm13[0,1],xmm11[2],xmm13[3,4],xmm11[5],xmm13[6,7]
+; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
+; AVX512F-NEXT: vpshufb %xmm0, %xmm10, %xmm10
; AVX512F-NEXT: vmovdqa64 %xmm0, %xmm26
-; AVX512F-NEXT: vinserti128 $1, %xmm8, %ymm12, %ymm8
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512F-NEXT: vinserti128 $1, %xmm11, %ymm10, %ymm10
+; AVX512F-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm10
; AVX512F-NEXT: vmovdqa 80(%rdi), %xmm12
-; AVX512F-NEXT: vmovdqa 80(%rsi), %xmm14
-; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512F-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
-; AVX512F-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512F-NEXT: vprold $16, %xmm14, %xmm14
+; AVX512F-NEXT: vmovdqa 80(%rsi), %xmm13
+; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512F-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,10,11,10,11,8,9,8,9,14,15,12,13,14,15]
+; AVX512F-NEXT: vpshufb %xmm11, %xmm15, %xmm15
+; AVX512F-NEXT: vprold $16, %xmm13, %xmm13
; AVX512F-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm14[2],xmm12[3,4],xmm14[5],xmm12[6,7]
-; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm12, %ymm0
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm8[4,5,6,7]
-; AVX512F-NEXT: vpermd %ymm7, %ymm20, %ymm8
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm21 = <5,5,u,6,6,u,7,7>
-; AVX512F-NEXT: vpermd %ymm15, %ymm21, %ymm12
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm12, %zmm12
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm16 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512F-NEXT: vpternlogq $184, %zmm0, %zmm16, %zmm12
-; AVX512F-NEXT: vmovdqa 96(%rdi), %ymm0
-; AVX512F-NEXT: vpshufb %ymm6, %ymm0, %ymm0
-; AVX512F-NEXT: vmovdqa 96(%rsi), %ymm8
-; AVX512F-NEXT: vmovdqa %ymm10, %ymm1
-; AVX512F-NEXT: vpshufb %ymm10, %ymm8, %ymm8
-; AVX512F-NEXT: vpor %ymm0, %ymm8, %ymm0
-; AVX512F-NEXT: vmovdqa 112(%rdi), %xmm8
-; AVX512F-NEXT: vmovdqa 112(%rsi), %xmm14
-; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm8[4],xmm14[5],xmm8[5],xmm14[6],xmm8[6],xmm14[7],xmm8[7]
-; AVX512F-NEXT: vpshufb %xmm13, %xmm10, %xmm10
-; AVX512F-NEXT: vprold $16, %xmm14, %xmm14
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm14[2],xmm8[3,4],xmm14[5],xmm8[6,7]
-; AVX512F-NEXT: vinserti128 $1, %xmm10, %ymm8, %ymm8
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm8[4,5,6,7]
-; AVX512F-NEXT: vpermd %ymm7, %ymm21, %ymm8
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm27 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512F-NEXT: vpandnq %ymm8, %ymm27, %ymm8
-; AVX512F-NEXT: vpshufb %ymm3, %ymm7, %ymm7
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm19
-; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm22 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512F-NEXT: vpternlogq $248, %zmm22, %zmm0, %zmm19
-; AVX512F-NEXT: vmovdqa 64(%rdi), %ymm0
-; AVX512F-NEXT: vpshufb %ymm6, %ymm0, %ymm0
-; AVX512F-NEXT: vmovdqa 64(%rsi), %ymm8
-; AVX512F-NEXT: vpshufb %ymm1, %ymm8, %ymm8
-; AVX512F-NEXT: vpor %ymm0, %ymm8, %ymm0
-; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-NEXT: vmovdqa 64(%rsi), %xmm8
-; AVX512F-NEXT: vprold $16, %xmm8, %xmm14
-; AVX512F-NEXT: vmovdqa 64(%rdi), %xmm7
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm9 = xmm7[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm9 = xmm9[0,1],xmm14[2],xmm9[3,4],xmm14[5],xmm9[6,7]
-; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
-; AVX512F-NEXT: vmovdqa64 %xmm26, %xmm10
-; AVX512F-NEXT: vpshufb %xmm10, %xmm7, %xmm7
-; AVX512F-NEXT: vinserti128 $1, %xmm9, %ymm7, %ymm7
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm7[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-NEXT: vpshufb %ymm3, %ymm15, %ymm7
-; AVX512F-NEXT: vpermd %ymm15, %ymm20, %ymm8
-; AVX512F-NEXT: vpandnq %ymm8, %ymm17, %ymm8
-; AVX512F-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
-; AVX512F-NEXT: vpternlogq $248, %zmm18, %zmm0, %zmm7
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm12 = xmm12[0,1],xmm13[2],xmm12[3,4],xmm13[5],xmm12[6,7]
+; AVX512F-NEXT: vinserti128 $1, %xmm15, %ymm12, %ymm12
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm12[0,1,2,3],zmm10[4,5,6,7]
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm20 = <5,5,u,6,6,u,7,7,u,8,8,u,9,9,u,10>
+; AVX512F-NEXT: vpermd 64(%rdx), %zmm20, %zmm10
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512F-NEXT: vpternlogq $184, %zmm15, %zmm21, %zmm10
+; AVX512F-NEXT: vmovdqa 96(%rdi), %ymm15
+; AVX512F-NEXT: vmovdqa %ymm7, %ymm1
+; AVX512F-NEXT: vpshufb %ymm7, %ymm15, %ymm15
+; AVX512F-NEXT: vmovdqa 96(%rsi), %ymm12
+; AVX512F-NEXT: vpshufb %ymm2, %ymm12, %ymm12
+; AVX512F-NEXT: vpor %ymm15, %ymm12, %ymm12
+; AVX512F-NEXT: vmovdqa 112(%rdi), %xmm15
+; AVX512F-NEXT: vmovdqa 112(%rsi), %xmm13
+; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
+; AVX512F-NEXT: vpshufb %xmm11, %xmm5, %xmm5
+; AVX512F-NEXT: vprold $16, %xmm13, %xmm13
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm13 = xmm15[0,1],xmm13[2],xmm15[3,4],xmm13[5],xmm15[6,7]
+; AVX512F-NEXT: vinserti128 $1, %xmm5, %ymm13, %ymm5
+; AVX512F-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm12[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-NEXT: vmovdqa 96(%rdx), %ymm12
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm23 = <5,5,u,6,6,u,7,7>
+; AVX512F-NEXT: vpermd %ymm12, %ymm23, %ymm15
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} ymm22 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512F-NEXT: vpandnq %ymm15, %ymm22, %ymm15
+; AVX512F-NEXT: vpshufb %ymm9, %ymm12, %ymm12
+; AVX512F-NEXT: vinserti64x4 $1, %ymm15, %zmm12, %zmm18
+; AVX512F-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512F-NEXT: vpternlogq $248, %zmm12, %zmm5, %zmm18
+; AVX512F-NEXT: vmovdqa 64(%rdi), %ymm5
+; AVX512F-NEXT: vpshufb %ymm7, %ymm5, %ymm5
+; AVX512F-NEXT: vmovdqa 64(%rsi), %ymm15
+; AVX512F-NEXT: vpshufb %ymm2, %ymm15, %ymm15
+; AVX512F-NEXT: vpor %ymm5, %ymm15, %ymm5
+; AVX512F-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512F-NEXT: vmovdqa 64(%rsi), %xmm15
+; AVX512F-NEXT: vprold $16, %xmm15, %xmm0
+; AVX512F-NEXT: vmovdqa 64(%rdi), %xmm13
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm7 = xmm13[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
+; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3]
+; AVX512F-NEXT: vmovdqa64 %xmm26, %xmm15
+; AVX512F-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm7, %ymm0
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-NEXT: vpshufb %ymm9, %ymm14, %ymm5
+; AVX512F-NEXT: vpermd %ymm14, %ymm19, %ymm7
+; AVX512F-NEXT: vpandnq %ymm7, %ymm16, %ymm7
+; AVX512F-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
+; AVX512F-NEXT: vpternlogq $248, %zmm17, %zmm0, %zmm5
; AVX512F-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX512F-NEXT: vpshufb %ymm6, %ymm0, %ymm0
-; AVX512F-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512F-NEXT: vpshufb %ymm1, %ymm8, %ymm8
-; AVX512F-NEXT: vpor %ymm0, %ymm8, %ymm0
-; AVX512F-NEXT: vmovdqa 48(%rdi), %xmm8
-; AVX512F-NEXT: vmovdqa 48(%rsi), %xmm9
-; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-NEXT: vpshufb %xmm13, %xmm14, %xmm14
-; AVX512F-NEXT: vprold $16, %xmm9, %xmm9
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0,1],xmm9[2],xmm8[3,4],xmm9[5],xmm8[6,7]
-; AVX512F-NEXT: vinserti128 $1, %xmm14, %ymm8, %ymm8
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
-; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm8[4,5,6,7]
-; AVX512F-NEXT: vpermd %ymm5, %ymm21, %ymm8
-; AVX512F-NEXT: vpandnq %ymm8, %ymm27, %ymm8
-; AVX512F-NEXT: vpshufb %ymm3, %ymm5, %ymm9
-; AVX512F-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512F-NEXT: vpternlogq $248, %zmm22, %zmm0, %zmm8
-; AVX512F-NEXT: vmovdqa64 %xmm25, %xmm1
-; AVX512F-NEXT: vprold $16, %xmm25, %xmm0
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0,1],xmm0[2],xmm9[3,4],xmm0[5],xmm9[6,7]
-; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm1[0],xmm11[1],xmm1[1],xmm11[2],xmm1[2],xmm11[3],xmm1[3]
-; AVX512F-NEXT: vpshufb %xmm10, %xmm3, %xmm3
-; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX512F-NEXT: vpshufb %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vmovdqa 32(%rsi), %ymm7
+; AVX512F-NEXT: vpshufb %ymm2, %ymm7, %ymm7
+; AVX512F-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512F-NEXT: vmovdqa 48(%rdi), %xmm7
+; AVX512F-NEXT: vmovdqa 48(%rsi), %xmm13
+; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm13[4],xmm7[4],xmm13[5],xmm7[5],xmm13[6],xmm7[6],xmm13[7],xmm7[7]
+; AVX512F-NEXT: vpshufb %xmm11, %xmm14, %xmm14
+; AVX512F-NEXT: vprold $16, %xmm13, %xmm13
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm13[2],xmm7[3,4],xmm13[5],xmm7[6,7]
+; AVX512F-NEXT: vinserti128 $1, %xmm14, %ymm7, %ymm7
+; AVX512F-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
+; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm7[4,5,6,7]
+; AVX512F-NEXT: vpermd %ymm8, %ymm23, %ymm7
+; AVX512F-NEXT: vpandnq %ymm7, %ymm22, %ymm7
+; AVX512F-NEXT: vpshufb %ymm9, %ymm8, %ymm8
+; AVX512F-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512F-NEXT: vpternlogq $248, %zmm12, %zmm0, %zmm7
+; AVX512F-NEXT: vprold $16, %xmm6, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm8 = xmm4[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0,1],xmm0[2],xmm8[3,4],xmm0[5],xmm8[6,7]
+; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
+; AVX512F-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512F-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-NEXT: vmovdqa64 %xmm23, %xmm6
; AVX512F-NEXT: vmovdqa64 %xmm24, %xmm1
-; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512F-NEXT: vpshufb %xmm13, %xmm3, %xmm3
-; AVX512F-NEXT: vprold $16, %xmm23, %xmm6
-; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm24[1,1,2,2]
-; AVX512F-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
-; AVX512F-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX512F-NEXT: vmovdqa64 %xmm25, %xmm6
+; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
+; AVX512F-NEXT: vpshufb %xmm11, %xmm2, %xmm2
+; AVX512F-NEXT: vprold $16, %xmm24, %xmm4
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm25[1,1,2,2]
+; AVX512F-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm4[2],xmm1[3,4],xmm4[5],xmm1[6,7]
+; AVX512F-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512F-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-NEXT: vpermd %ymm5, %ymm20, %ymm1
-; AVX512F-NEXT: vpermd %ymm2, %ymm21, %ymm2
-; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512F-NEXT: vpternlogq $184, %zmm0, %zmm16, %zmm1
+; AVX512F-NEXT: vpermd (%rdx), %zmm20, %zmm1
+; AVX512F-NEXT: vpternlogq $184, %zmm0, %zmm21, %zmm1
; AVX512F-NEXT: vmovdqa64 %zmm1, 64(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm8, 128(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm7, 192(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm19, 320(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm12, 256(%rcx)
-; AVX512F-NEXT: vmovdqa64 %zmm4, (%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm7, 128(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm5, 192(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm18, 320(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm10, 256(%rcx)
+; AVX512F-NEXT: vmovdqa64 %zmm3, (%rcx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
index e2ca0de7b33f..264655bb0ce8 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
@@ -3539,242 +3539,238 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf32:
; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $24, %rsp
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm13
+; AVX512F-ONLY-FAST-NEXT: pushq %rax
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm13[0],ymm9[1],ymm13[1],ymm9[2],ymm13[2],ymm9[3],ymm13[3],ymm9[8],ymm13[8],ymm9[9],ymm13[9],ymm9[10],ymm13[10],ymm9[11],ymm13[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [5,4,2,2,5,4,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm1[0],ymm9[1],ymm1[1],ymm9[2],ymm1[2],ymm9[3],ymm1[3],ymm9[8],ymm1[8],ymm9[9],ymm1[9],ymm9[10],ymm1[10],ymm9[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm25
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = <1,2,1,2,u,u,3,3>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm24, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm23
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm1, %zmm23, %zmm21
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm15
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm15[0],ymm4[0],ymm15[1],ymm4[1],ymm15[2],ymm4[2],ymm15[3],ymm4[3],ymm15[8],ymm4[8],ymm15[9],ymm4[9],ymm15[10],ymm4[10],ymm15[11],ymm4[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-FAST-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [8,9,20,11,12,21,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm23, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm21, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm7, %xmm1, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm27
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [0,9,2,3,8,5,6,11]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm17, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm1, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm21 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [8,9,20,11,12,21,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm19, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm14[0],ymm1[0],ymm14[1],ymm1[1],ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[8],ymm1[8],ymm14[9],ymm1[9],ymm14[10],ymm1[10],ymm14[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm6, %xmm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm5, %xmm16
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [0,9,2,3,8,5,6,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm20, %ymm21
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm0, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm5, %xmm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm29
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm3[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm6, %xmm16
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm14[0],ymm2[1],ymm14[1],ymm2[2],ymm14[2],ymm2[3],ymm14[3],ymm2[8],ymm14[8],ymm2[9],ymm14[9],ymm2[10],ymm14[10],ymm2[11],ymm14[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm3, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm24, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm12, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm6, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm12, %zmm24 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm6 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm6, %zmm24, %zmm21
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm5 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm5[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm10[1,1,1,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm6, %zmm23, %zmm24
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm22, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm24 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm6 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm6, %zmm24, %zmm19
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm7, %xmm6, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm17, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm7, %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm15[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [5,6,5,6,5,6,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm15 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm22, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm6, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm20, %ymm24
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm14[4],ymm1[4],ymm14[5],ymm1[5],ymm14[6],ymm1[6],ymm14[7],ymm1[7],ymm14[12],ymm1[12],ymm14[13],ymm1[13],ymm14[14],ymm1[14],ymm14[15],ymm1[15]
; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm14, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[8],ymm0[8],ymm11[9],ymm0[9],ymm11[10],ymm0[10],ymm11[11],ymm0[11]
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm14 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm14[0],ymm1[0],ymm14[1],ymm1[1],ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[8],ymm1[8],ymm14[9],ymm1[9],ymm14[10],ymm1[10],ymm14[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm10, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm0[0],ymm13[1],ymm0[1],ymm13[2],ymm0[2],ymm13[3],ymm0[3],ymm13[8],ymm0[8],ymm13[9],ymm0[9],ymm13[10],ymm0[10],ymm13[11],ymm0[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [5,6,5,6,5,6,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm13, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm2[4],ymm14[4],ymm2[5],ymm14[5],ymm2[6],ymm14[6],ymm2[7],ymm14[7],ymm2[12],ymm14[12],ymm2[13],ymm14[13],ymm2[14],ymm14[14],ymm2[15],ymm14[15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm14 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [0,13,2,3,12,5,6,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm25, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm11, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm9[4],ymm13[4],ymm9[5],ymm13[5],ymm9[6],ymm13[6],ymm9[7],ymm13[7],ymm9[12],ymm13[12],ymm9[13],ymm13[13],ymm9[14],ymm13[14],ymm9[15],ymm13[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm22, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm12[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm12 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm10 = ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm14 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm20
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [0,13,2,3,12,5,6,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm23, %ymm20
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [12,1,2,13,4,5,14,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm3, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm25, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm2[4],ymm9[5],ymm2[5],ymm9[6],ymm2[6],ymm9[7],ymm2[7],ymm9[12],ymm2[12],ymm9[13],ymm2[13],ymm9[14],ymm2[14],ymm9[15],ymm2[15]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm13, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm26, %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm0, %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm11[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm11 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm12[0],ymm9[1],ymm12[1],ymm9[2],ymm12[2],ymm9[3],ymm12[3],ymm9[8],ymm12[8],ymm9[9],ymm12[9],ymm9[10],ymm12[10],ymm9[11],ymm12[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm28, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm13, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm26, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm12[0],ymm3[1],ymm12[1],ymm3[2],ymm12[2],ymm3[3],ymm12[3],ymm3[8],ymm12[8],ymm3[9],ymm12[9],ymm3[10],ymm12[10],ymm3[11],ymm12[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm13 = ymm0[4],ymm13[4],ymm0[5],ymm13[5],ymm0[6],ymm13[6],ymm0[7],ymm13[7],ymm0[12],ymm13[12],ymm0[13],ymm13[13],ymm0[14],ymm13[14],ymm0[15],ymm13[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm15, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm0[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm9[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm13[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm9 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm10, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm25, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm15, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm1[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm11, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm8, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm1[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm0[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm11[0],ymm9[1],ymm11[1],ymm9[2],ymm11[2],ymm9[3],ymm11[3],ymm9[8],ymm11[8],ymm9[9],ymm11[9],ymm9[10],ymm11[10],ymm9[11],ymm11[11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm15, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm11[0],ymm1[1],ymm11[1],ymm1[2],ymm11[2],ymm1[3],ymm11[3],ymm1[8],ymm11[8],ymm1[9],ymm11[9],ymm1[10],ymm11[10],ymm1[11],ymm11[11]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm15[4],ymm4[4],ymm15[5],ymm4[5],ymm15[6],ymm4[6],ymm15[7],ymm4[7],ymm15[12],ymm4[12],ymm15[13],ymm4[13],ymm15[14],ymm4[14],ymm15[15],ymm4[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm15 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm15[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm9, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm9, %zmm4 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm4, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm12, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm23, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm2[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm10[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm10 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm10, %ymm3, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm1[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm1[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm16, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm17, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm8, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm5, %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7]
; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [1,0,2,2,1,0,2,2]
; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm11, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm2 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm16, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm5, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm18, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm8, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm4[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm11, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm5[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm2
; AVX512F-ONLY-FAST-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [16,9,10,17,12,13,18,15]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm2, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm3, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm31, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm2 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [16,9,10,17,12,13,18,15]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm5, %zmm3, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm11, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm12 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm11, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm8, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm15, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm12[4],xmm1[5],xmm12[5],xmm1[6],xmm12[6],xmm1[7],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm1, %xmm15
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm10, %xmm13
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm12[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm7, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm27, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm8[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm7, %zmm1, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm15[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm7, %zmm11 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm7, %zmm11, %zmm3
; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm27[0],zero,xmm27[1],zero,xmm27[2],zero,xmm27[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm8[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm11, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm8, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm13, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm23[0,1,2,3],zmm20[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm16 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm12, %zmm18, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm24[0,1,2,3],zmm21[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm19, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm12, %zmm18, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm14[0,1,2,3],zmm12[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm26, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm15 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm12, %zmm15, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm9[0,1,2,3],zmm12[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm25, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm15, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,8,3,4,9,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm9, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm10[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm21[0,1,2,3],zmm18[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm1, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm15, %zmm17, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm24[0,1,2,3],zmm19[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm22, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm15, %zmm17, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm15
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm14[0,1,2,3],zmm15[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm26, %zmm15
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm14, %zmm17, %zmm15
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm28, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm4, %zmm17, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,8,3,4,9,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm5, %ymm4, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm12[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm9, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm2[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm5, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm4, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm11[0,1,2,3],zmm3[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm13, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm5, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, 320(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 256(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $24, %rsp
+; AVX512F-ONLY-FAST-NEXT: popq %rax
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
@@ -4033,233 +4029,226 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf32:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm26
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm24
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [5,4,2,2,5,4,6,6]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm5, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <1,2,1,2,u,u,3,3>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm21
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm15[0],ymm4[0],ymm15[1],ymm4[1],ymm15[2],ymm4[2],ymm15[3],ymm4[3],ymm15[8],ymm4[8],ymm15[9],ymm4[9],ymm15[10],ymm4[10],ymm15[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm21 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm12
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm18
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm19, %zmm18
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm1, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm17
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm20, %ymm21
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm2, %xmm27
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm28
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm6
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm0[2,2,2,2]
+; AVX512DQ-FAST-NEXT: subq $40, %rsp
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm2
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm22
; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm5, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm16
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm0[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[2],ymm8[2],ymm1[3],ymm8[3],ymm1[8],ymm8[8],ymm1[9],ymm8[9],ymm1[10],ymm8[10],ymm1[11],ymm8[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm11, %zmm25
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm22, %zmm25 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm11
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm7, %zmm25, %zmm19
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm9
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm9, %xmm7
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm7, %ymm20, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm8, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm0[0],ymm13[1],ymm0[1],ymm13[2],ymm0[2],ymm13[3],ymm0[3],ymm13[8],ymm0[8],ymm13[9],ymm0[9],ymm13[10],ymm0[10],ymm13[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm8[4],ymm1[5],ymm8[5],ymm1[6],ymm8[6],ymm1[7],ymm8[7],ymm1[12],ymm8[12],ymm1[13],ymm8[13],ymm1[14],ymm8[14],ymm1[15],ymm8[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm14[4],ymm2[4],ymm14[5],ymm2[5],ymm14[6],ymm2[6],ymm14[7],ymm2[7],ymm14[12],ymm2[12],ymm14[13],ymm2[13],ymm14[14],ymm2[14],ymm14[15],ymm2[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm13, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm4, %ymm1
; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm14 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm3 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm29
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm14
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm2, %zmm14 {%k1}
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm20
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [0,13,2,3,12,5,6,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm11, %ymm0
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm23, %ymm20
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm2, %ymm14
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm4, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm15, %ymm7
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[2],ymm0[2],ymm7[3],ymm0[3],ymm7[8],ymm0[8],ymm7[9],ymm0[9],ymm7[10],ymm0[10],ymm7[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm27, %xmm1
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm11
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm11[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm15[4],ymm4[4],ymm15[5],ymm4[5],ymm15[6],ymm4[6],ymm15[7],ymm4[7],ymm15[12],ymm4[12],ymm15[13],ymm4[13],ymm15[14],ymm4[14],ymm15[15],ymm4[15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm26, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm24, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm11 = ymm15[4],ymm1[4],ymm15[5],ymm1[5],ymm15[6],ymm1[6],ymm15[7],ymm1[7],ymm15[12],ymm1[12],ymm15[13],ymm1[13],ymm15[14],ymm1[14],ymm15[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm13, %ymm11
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm15 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm1[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm7[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm13[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm11, %zmm7, %zmm4 {%k1}
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm4, %ymm13
-; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm12, %ymm1
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm23, %ymm13
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm15, %ymm7
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm7[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm15, %ymm7
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm7[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm7, %ymm2, %ymm4
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm6, %ymm1
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm27 = ymm1[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm6, %ymm1
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm1[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm5, %xmm1
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm3, %xmm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm16, %xmm7
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm10[0],xmm7[1],xmm10[1],xmm7[2],xmm10[2],xmm7[3],xmm10[3]
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FAST-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm5, %ymm3
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm6, %ymm2
-; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm6
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm2, %zmm6
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm9 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm9, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm29, %xmm8
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm12
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm5, %ymm5
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm15, %zmm10
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm5, %zmm10 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm7
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm7[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm5, %zmm10, %zmm2
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm17[0],zero,xmm17[1],zero,xmm17[2],zero,xmm17[3],zero
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm28, %xmm8
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm8[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm11
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm2, %zmm22 {%k1}
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm22, %ymm20
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [0,13,2,3,12,5,6,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm2, %ymm8, %ymm20
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm2, %ymm11, %ymm22
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm3
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm17
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm3[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm5, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm15
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm15, %ymm0
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[8],ymm3[8],ymm0[9],ymm3[9],ymm0[10],ymm3[10],ymm0[11],ymm3[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm15[4],ymm5[4],ymm15[5],ymm5[5],ymm15[6],ymm5[6],ymm15[7],ymm5[7],ymm15[12],ymm5[12],ymm15[13],ymm5[13],ymm15[14],ymm5[14],ymm15[15],ymm5[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm16
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm23
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm13
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm13[4],ymm3[5],ymm13[5],ymm3[6],ymm13[6],ymm3[7],ymm13[7],ymm3[12],ymm13[12],ymm3[13],ymm13[13],ymm3[14],ymm13[14],ymm3[15],ymm13[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm9 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm9[0],ymm4[0],ymm9[1],ymm4[1],ymm9[2],ymm4[2],ymm9[3],ymm4[3],ymm9[8],ymm4[8],ymm9[9],ymm4[9],ymm9[10],ymm4[10],ymm9[11],ymm4[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm4, %zmm23 {%k1}
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm23, %ymm21
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm9
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm8, %ymm21
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm11, %ymm23
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm8
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm8, %ymm0
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm8, %ymm0
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm12
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm6
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm13[0],ymm3[1],ymm13[1],ymm3[2],ymm13[2],ymm3[3],ymm13[3],ymm3[8],ymm13[8],ymm3[9],ymm13[9],ymm3[10],ymm13[10],ymm3[11],ymm13[11]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm11[4],xmm6[5],xmm11[5],xmm6[6],xmm11[6],xmm6[7],xmm11[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm25
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm12, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm3
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm0[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm0[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FAST-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm13
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm26, %ymm5
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm0 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm3
+; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512DQ-FAST-NEXT: kmovw %eax, %k2
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm5, %zmm3 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm27
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm10
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm10[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm24, %zmm27
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm0
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm7, %xmm2
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm26, %ymm26
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm14, %ymm28
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm0
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm26, %zmm0 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm6
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm11, %zmm0, %zmm24
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm7
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm15[0],ymm7[0],ymm15[1],ymm7[1],ymm15[2],ymm7[2],ymm15[3],ymm7[3],ymm15[8],ymm7[8],ymm15[9],ymm7[9],ymm15[10],ymm7[10],ymm15[11],ymm7[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512DQ-FAST-NEXT: vpermd %zmm25, %zmm11, %zmm15
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm5, %zmm15 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, %zmm9
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm7, %zmm5, %zmm9
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm29, %ymm14
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm7 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm7 = ymm14[0],mem[0],ymm14[1],mem[1],ymm14[2],mem[2],ymm14[3],mem[3],ymm14[8],mem[8],ymm14[9],mem[9],ymm14[10],mem[10],ymm14[11],mem[11]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm6, %xmm14
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm14, %ymm25, %ymm15
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm2, %xmm14
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm14[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512DQ-FAST-NEXT: vpermd %zmm1, %zmm11, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm11
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm11[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm12, %ymm29
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm11, %xmm13
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm14 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm14 = ymm12[0],mem[0],ymm12[1],mem[1],ymm12[2],mem[2],ymm12[3],mem[3],ymm12[8],mem[8],ymm12[9],mem[9],ymm12[10],mem[10],ymm12[11],mem[11]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm11, %xmm11
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm14, %zmm4, %zmm1 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm4
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm4, %zmm1, %zmm5
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm10, %xmm7
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm7, %ymm25, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm17, %ymm7
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm10
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm22[0,1,2,3],zmm10[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm12, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm10, %zmm14, %zmm12
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm10
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm23[0,1,2,3],zmm10[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm16
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm10, %zmm14, %zmm16
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm10, %ymm3
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm10, %ymm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm27[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm28, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm6, %zmm4
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm24[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm29, %zmm3
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm6, %zmm3
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm15[0,1,2,3],zmm9[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm6
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm5[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm1
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm1
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm21[0,1,2,3],zmm18[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm8, %zmm15
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm12, %zmm16, %zmm15
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm25[0,1,2,3],zmm19[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm22, %zmm17
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm12, %zmm16, %zmm17
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm12
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm14[0,1,2,3],zmm12[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm26, %zmm14
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm12, %zmm16, %zmm14
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm12
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm12[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm27, %zmm8
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm16, %zmm8
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm4, %ymm1
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm6[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm3, %zmm0
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm4, %ymm10
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm10[0,1,2,3],zmm2[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm3, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, 64(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 64(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, 320(%rax)
+; AVX512DQ-FAST-NEXT: addq $40, %rsp
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
@@ -7992,492 +7981,489 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf64:
; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $1256, %rsp # imm = 0x4E8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-ONLY-FAST-NEXT: subq $936, %rsp # imm = 0x3A8
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [5,4,2,2,5,4,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 64(%rcx), %xmm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm7, (%rsp) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm8, %xmm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm2, %xmm29
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm14[0],ymm9[1],ymm14[1],ymm9[2],ymm14[2],ymm9[3],ymm14[3],ymm9[8],ymm14[8],ymm9[9],ymm14[9],ymm9[10],ymm14[10],ymm9[11],ymm14[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm26, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm5, %xmm25
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm31
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm15
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm28
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm3
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = <1,2,1,2,u,u,3,3>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm28, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm3
-; AVX512F-ONLY-FAST-NEXT: movw $18724, %ax # imm = 0x4924
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [8,9,20,11,12,21,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm0[1,1,1,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm0[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm1, %zmm20, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm1
+; AVX512F-ONLY-FAST-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm18 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [8,9,20,11,12,21,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm21, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [0,9,2,3,8,5,6,11]
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, %xmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [0,9,2,3,8,5,6,11]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm21, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm18
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, %xmm10
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm1[0,1,0,1]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm1[2,2,2,2]
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm7[0],ymm5[0],ymm7[1],ymm5[1],ymm7[2],ymm5[2],ymm7[3],ymm5[3],ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm16, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm28, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm12, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm19 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm24, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm15, %xmm8, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm21, %ymm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm8 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = [5,6,5,6,5,6,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm7[4],ymm5[4],ymm7[5],ymm5[5],ymm7[6],ymm5[6],ymm7[7],ymm5[7],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm18, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm7 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm7[0],ymm5[0],ymm7[1],ymm5[1],ymm7[2],ymm5[2],ymm7[3],ymm5[3],ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm12[0],ymm7[0],ymm12[1],ymm7[1],ymm12[2],ymm7[2],ymm12[3],ymm7[3],ymm12[8],ymm7[8],ymm12[9],ymm7[9],ymm12[10],ymm7[10],ymm12[11],ymm7[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm0[0],ymm8[0],ymm0[1],ymm8[1],ymm0[2],ymm8[2],ymm0[3],ymm8[3],ymm0[8],ymm8[8],ymm0[9],ymm8[9],ymm0[10],ymm8[10],ymm0[11],ymm8[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm16, %zmm20, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm16 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm21, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [5,6,5,6,5,6,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm12[4],ymm11[4],ymm12[5],ymm11[5],ymm12[6],ymm11[6],ymm12[7],ymm11[7],ymm12[12],ymm11[12],ymm12[13],ymm11[13],ymm12[14],ymm11[14],ymm12[15],ymm11[15]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm23, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm3 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm11 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm11[0],ymm3[0],ymm11[1],ymm3[1],ymm11[2],ymm3[2],ymm11[3],ymm3[3],ymm11[8],ymm3[8],ymm11[9],ymm3[9],ymm11[10],ymm3[10],ymm11[11],ymm3[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm3[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm8, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[8],ymm12[8],ymm10[9],ymm12[9],ymm10[10],ymm12[10],ymm10[11],ymm12[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm8[4],ymm0[5],ymm8[5],ymm0[6],ymm8[6],ymm0[7],ymm8[7],ymm0[12],ymm8[12],ymm0[13],ymm8[13],ymm0[14],ymm8[14],ymm0[15],ymm8[15]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [0,13,2,3,12,5,6,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm5, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm5, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm1, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm11, %zmm8
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm8, %zmm12 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm12, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [0,13,2,3,12,5,6,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm3, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm3, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [12,1,2,13,4,5,14,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm3, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm5[4],ymm1[5],ymm5[5],ymm1[6],ymm5[6],ymm1[7],ymm5[7],ymm1[12],ymm5[12],ymm1[13],ymm5[13],ymm1[14],ymm5[14],ymm1[15],ymm5[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm12[0],ymm8[0],ymm12[1],ymm8[1],ymm12[2],ymm8[2],ymm12[3],ymm8[3],ymm12[8],ymm8[8],ymm12[9],ymm8[9],ymm12[10],ymm8[10],ymm12[11],ymm8[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm12 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm8, %zmm22
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm22 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm22, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm3, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm9, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm17, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm20, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm20 = [1,0,2,2,1,0,2,2]
-; AVX512F-ONLY-FAST-NEXT: # ymm20 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm20, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm4, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm4, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm14[4],ymm9[5],ymm14[5],ymm9[6],ymm14[6],ymm9[7],ymm14[7],ymm9[12],ymm14[12],ymm9[13],ymm14[13],ymm9[14],ymm14[14],ymm9[15],ymm14[15]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm23, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm6 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[8],ymm1[8],ymm6[9],ymm1[9],ymm6[10],ymm1[10],ymm6[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm8, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm11, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[8],ymm6[8],ymm9[9],ymm6[9],ymm9[10],ymm6[10],ymm9[11],ymm6[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm11[4],ymm8[4],ymm11[5],ymm8[5],ymm11[6],ymm8[6],ymm11[7],ymm8[7],ymm11[12],ymm8[12],ymm11[13],ymm8[13],ymm11[14],ymm8[14],ymm11[15],ymm8[15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm14 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm0, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm10, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm3, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm30 = [1,0,2,2,1,0,2,2]
+; AVX512F-ONLY-FAST-NEXT: # ymm30 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm17, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm30, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm10
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm29
; AVX512F-ONLY-FAST-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm6, %zmm30 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm31 = [16,9,10,17,12,13,18,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm30, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm13[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm31, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm20, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm7, %ymm11
+; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm29 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm29, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm17, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm26, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm25, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm30, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm7, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm8, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm29
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm10, %zmm29 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm29, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm11[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm10, %zmm31, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[1],ymm5[1],ymm1[2],ymm5[2],ymm1[3],ymm5[3],ymm1[8],ymm5[8],ymm1[9],ymm5[9],ymm1[10],ymm5[10],ymm1[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm28, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm26
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm26 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm23
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm24, %zmm23
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm15, %xmm11, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm21, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm19, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm5, %zmm26 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm26, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm2, %zmm17, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm8[0],ymm11[1],ymm8[1],ymm11[2],ymm8[2],ymm11[3],ymm8[3],ymm11[8],ymm8[8],ymm11[9],ymm8[9],ymm11[10],ymm8[10],ymm11[11],ymm8[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm31, %zmm20, %zmm31
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm31 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm31, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm21, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm22, %ymm31
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm28, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[8],ymm0[8],ymm4[9],ymm0[9],ymm4[10],ymm0[10],ymm4[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm28 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm10, %zmm28, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm15, %xmm8, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm10, %ymm21, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm4, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[8],ymm6[8],ymm10[9],ymm6[9],ymm10[10],ymm6[10],ymm10[11],ymm6[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm13[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[8],ymm5[8],ymm0[9],ymm5[9],ymm0[10],ymm5[10],ymm0[11],ymm5[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm28, %zmm20, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm20 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm1, %zmm20, %zmm21
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm20
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[12],ymm13[12],ymm15[13],ymm13[13],ymm15[14],ymm13[14],ymm15[15],ymm13[15]
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm13[0],ymm4[1],ymm13[1],ymm4[2],ymm13[2],ymm4[3],ymm13[3],ymm4[8],ymm13[8],ymm4[9],ymm13[9],ymm4[10],ymm13[10],ymm4[11],ymm13[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm15
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm15
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm11[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[12],ymm0[12],ymm4[13],ymm0[13],ymm4[14],ymm0[14],ymm4[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm16, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm4[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm18, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm4, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm3, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm1, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [0,13,2,3,12,5,6,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm5, %ymm15, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm5[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm5 = ymm10[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm5[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm0[4],ymm5[4],ymm0[5],ymm5[5],ymm0[6],ymm5[6],ymm0[7],ymm5[7],ymm0[12],ymm5[12],ymm0[13],ymm5[13],ymm0[14],ymm5[14],ymm0[15],ymm5[15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm8, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm23, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm13, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm4 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm4, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [0,13,2,3,12,5,6,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm10[4],ymm0[5],ymm10[5],ymm0[6],ymm10[6],ymm0[7],ymm10[7],ymm0[12],ymm10[12],ymm0[13],ymm10[13],ymm0[14],ymm10[14],ymm0[15],ymm10[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm5, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm9[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm9 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm10 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm12, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm13[0],ymm10[0],ymm13[1],ymm10[1],ymm13[2],ymm10[2],ymm13[3],ymm10[3],ymm13[8],ymm10[8],ymm13[9],ymm10[9],ymm13[10],ymm10[10],ymm13[11],ymm10[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm14 = ymm0[4],ymm12[4],ymm0[5],ymm12[5],ymm0[6],ymm12[6],ymm0[7],ymm12[7],ymm0[12],ymm12[12],ymm0[13],ymm12[13],ymm0[14],ymm12[14],ymm0[15],ymm12[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm0[4],ymm6[5],ymm0[5],ymm6[6],ymm0[6],ymm6[7],ymm0[7],ymm6[12],ymm0[12],ymm6[13],ymm0[13],ymm6[14],ymm0[14],ymm6[15],ymm0[15]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm23, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm11 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm11[0],ymm13[0],ymm11[1],ymm13[1],ymm11[2],ymm13[2],ymm11[3],ymm13[3],ymm11[8],ymm13[8],ymm11[9],ymm13[9],ymm11[10],ymm13[10],ymm11[11],ymm13[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm24, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm9, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm13[0],ymm3[1],ymm13[1],ymm3[2],ymm13[2],ymm3[3],ymm13[3],ymm3[8],ymm13[8],ymm3[9],ymm13[9],ymm3[10],ymm13[10],ymm3[11],ymm13[11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm13
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm13[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm13 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm13[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm15 = ymm9[4],ymm6[4],ymm9[5],ymm6[5],ymm9[6],ymm6[6],ymm9[7],ymm6[7],ymm9[12],ymm6[12],ymm9[13],ymm6[13],ymm9[14],ymm6[14],ymm9[15],ymm6[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm6, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm9[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm11, %zmm1 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm6, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm1, %ymm23
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm22, %ymm23
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm13 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm13[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm14 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm10, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm9, %ymm15, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm9, %ymm8, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm17, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm16, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm0, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm11 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [12,1,2,13,4,5,14,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm7, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm6, %ymm27
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm1, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm8, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm13[4],xmm7[4],xmm13[5],xmm7[5],xmm13[6],xmm7[6],xmm13[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm2 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsp), %xmm9 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm30, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm0 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm6, %xmm10, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm6, %xmm8, %xmm15
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm7, %zmm31, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm20, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm13 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm13 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm13 = zmm11[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm15 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm15
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm13 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm13 = zmm19[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm19 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm26[0,1,2,3],zmm23[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm17 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm13 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm13 = zmm28[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm21, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm11[0,1,2,3],zmm13[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm18 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm18
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm22[0,1,2,3],zmm13[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm21 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm13, %zmm20, %zmm21
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm9[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm10[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm4, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm3, %zmm20, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # xmm3 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm13[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm7, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm30, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm9 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # zmm9 = zmm18[0,1,2,3],mem[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm15 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm15
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm9 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # zmm9 = zmm16[0,1,2,3],mem[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm16 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm31[0,1,2,3],zmm19[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm28, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm20[0,1,2,3],zmm21[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm24, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm12[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm12 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm14[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm14 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm9, %zmm18, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm27[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm22[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm4, %zmm18, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm20 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm20 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm11, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm4, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm9
; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm9 = xmm12[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm9, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm12, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm9 = xmm10[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm9, %ymm21
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm10
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm22 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # xmm22 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm10, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm6, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm25[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm25 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm7, %zmm8, %zmm31
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm24 = mem[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq $230, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm25 = mem[2,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm27 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm7, %zmm4 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm7, %zmm4, %zmm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm13[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm5, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm26 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsp), %xmm11 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm11, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm11, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm13, %xmm13
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm14[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm23, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm20, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm28 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm0, %zmm23
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm23[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm23
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm18, %zmm23
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,8,3,4,9,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm1, %ymm30
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm22, %ymm1, %ymm29
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm25, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm26, %ymm1, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm30[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm20, %ymm1, %ymm29
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm22, %ymm1, %ymm26
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm27, %ymm1, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm28, %ymm1, %ymm4
; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm1 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm29[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm21, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm18, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm1 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm26[0,1,2,3],mem[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm18, %zmm3
; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm0[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm9, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm31[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm9, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm18, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm17[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm18, %zmm2
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, (%rax)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, 448(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 576(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm21, 512(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 704(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, 640(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm23, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm11, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, 448(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 384(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 576(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 512(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, 704(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 640(%rax)
; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $1256, %rsp # imm = 0x4E8
+; AVX512F-ONLY-FAST-NEXT: addq $936, %rsp # imm = 0x3A8
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
@@ -8979,481 +8965,467 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf64:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $1144, %rsp # imm = 0x478
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm0
+; AVX512DQ-FAST-NEXT: subq $936, %rsp # imm = 0x3A8
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm12
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 96(%rcx), %xmm19
+; AVX512DQ-FAST-NEXT: vmovdqa64 96(%rdx), %xmm17
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm13
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovaps 32(%rdx), %xmm0
+; AVX512DQ-FAST-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm19, %xmm5
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm6[0],ymm12[0],ymm6[1],ymm12[1],ymm6[2],ymm12[2],ymm6[3],ymm12[3],ymm6[8],ymm12[8],ymm6[9],ymm12[9],ymm6[10],ymm12[10],ymm6[11],ymm12[11]
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, %xmm11
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm15
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm15[0],ymm14[0],ymm15[1],ymm14[1],ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[8],ymm14[8],ymm15[9],ymm14[9],ymm15[10],ymm14[10],ymm15[11],ymm14[11]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm9, %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [5,4,2,2,5,4,6,6]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = <1,2,1,2,u,u,3,3>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm27, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm22
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm8
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm0[4],ymm9[4],ymm0[5],ymm9[5],ymm0[6],ymm9[6],ymm0[7],ymm9[7],ymm0[12],ymm9[12],ymm0[13],ymm9[13],ymm0[14],ymm9[14],ymm0[15],ymm9[15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm31 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm31, %ymm9
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[8],ymm0[8],ymm4[9],ymm0[9],ymm4[10],ymm0[10],ymm4[11],ymm0[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm20
; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm24 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm1
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm2, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, %zmm21
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm17, %ymm24
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, %xmm13
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm9, %zmm0, %zmm20 {%k1}
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm20, %ymm21
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [0,13,2,3,12,5,6,15]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm23, %ymm21
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm22, %ymm20
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm25, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm7
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm27, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm4[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm10
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm10[0],ymm4[0],ymm10[1],ymm4[1],ymm10[2],ymm4[2],ymm10[3],ymm4[3],ymm10[8],ymm4[8],ymm10[9],ymm4[9],ymm10[10],ymm4[10],ymm10[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm9
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm12, %zmm11, %zmm9 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm11
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm2
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm21, %zmm11
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm11
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm11, %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm11, %xmm15
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm17, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, (%rsp) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm3, %xmm3
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm3
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm11 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm4, %ymm12
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm10, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm10[4],ymm4[4],ymm10[5],ymm4[5],ymm10[6],ymm4[6],ymm10[7],ymm4[7],ymm10[12],ymm4[12],ymm10[13],ymm4[13],ymm10[14],ymm4[14],ymm10[15],ymm4[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm10 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm20, %ymm10
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm12, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm0, %zmm4 {%k1}
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm4, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [0,13,2,3,12,5,6,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm0
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm19, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm1, %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm4, %ymm7
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[2],ymm0[2],ymm7[3],ymm0[3],ymm7[8],ymm0[8],ymm7[9],ymm0[9],ymm7[10],ymm0[10],ymm7[11],ymm0[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm6[4],ymm12[4],ymm6[5],ymm12[5],ymm6[6],ymm12[6],ymm6[7],ymm12[7],ymm6[12],ymm12[12],ymm6[13],ymm12[13],ymm6[14],ymm12[14],ymm6[15],ymm12[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm31, %ymm8
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm9 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[2],ymm9[2],ymm6[3],ymm9[3],ymm6[8],ymm9[8],ymm6[9],ymm9[9],ymm6[10],ymm9[10],ymm6[11],ymm9[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm27
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm6, %zmm27 {%k1}
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm27, %ymm28
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm29
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm23, %ymm28
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm22, %ymm27
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm30
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm18
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm4
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm4, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm3
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm0[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm12 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm20, %ymm12
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm14 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm28
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm12, %zmm13, %zmm28 {%k1}
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm28, %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm1
-; AVX512DQ-FAST-NEXT: vpshufb %ymm9, %ymm1, %ymm10
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm10, %ymm19, %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm10, %ymm16, %ymm28
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm6, %xmm10
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm8, %xmm12
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FAST-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm12, %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa %ymm12, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm30
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm3, %xmm7
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FAST-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm19, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm4
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm16, %ymm8
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm24
; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FAST-NEXT: kmovw %eax, %k2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm8, %zmm30 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm31 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm30, %zmm6
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm15[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm31, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm5, %xmm7
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm6, %xmm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm7[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm7[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm8
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm9, %ymm14
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm13 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm13, %ymm15
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm16
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm14, %zmm16 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm9
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm10, %zmm31, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm27, %ymm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm3, %zmm5, %zmm26 {%k1}
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm26, %zmm18
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm21, %zmm18
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm9, %xmm1
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm17, %ymm26
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm25, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm27, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm6[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm15[0],ymm6[0],ymm15[1],ymm6[1],ymm15[2],ymm6[2],ymm15[3],ymm6[3],ymm15[8],ymm6[8],ymm15[9],ymm6[9],ymm15[10],ymm6[10],ymm15[11],ymm6[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm27
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm14, %zmm12, %zmm27 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm10 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm10, %zmm27, %zmm21
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm9, %xmm10
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm10, %ymm17, %ymm27
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm6, %ymm9
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm15, %ymm10
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25>
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm9, %ymm12
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm12[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm15[4],ymm6[4],ymm15[5],ymm6[5],ymm15[6],ymm6[6],ymm15[7],ymm6[7],ymm15[12],ymm6[12],ymm15[13],ymm6[13],ymm15[14],ymm6[14],ymm15[15],ymm6[15]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm12 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm23, %ymm17
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm23, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm9, %ymm2
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm2[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm3, %zmm8, %zmm24 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm25
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm17, %zmm25
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm1, %xmm5
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm13[0],xmm11[1],xmm13[1],xmm11[2],xmm13[2],xmm11[3],xmm13[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm16, %ymm8
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm13 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm13[0],xmm11[1],xmm13[1],xmm11[2],xmm13[2],xmm11[3],xmm13[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm11, %ymm13
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm5, %zmm11
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm13, %zmm8, %zmm11 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm11, %zmm26
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm17, %zmm26
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm10[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm20, %ymm10
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm25
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm1, %zmm25 {%k1}
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm25, %ymm13
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm19, %ymm14
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm19, %ymm13
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm3, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm22, %ymm6
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm6, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm4, %ymm5
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm3[0],ymm5[1],ymm3[1],ymm5[2],ymm3[2],ymm5[3],ymm3[3],ymm5[8],ymm3[8],ymm5[9],ymm3[9],ymm5[10],ymm3[10],ymm5[11],ymm3[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[12],ymm0[12],ymm4[13],ymm0[13],ymm4[14],ymm0[14],ymm4[15],ymm0[15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm15[4],ymm14[4],ymm15[5],ymm14[5],ymm15[6],ymm14[6],ymm15[7],ymm14[7],ymm15[12],ymm14[12],ymm15[13],ymm14[13],ymm15[14],ymm14[14],ymm15[15],ymm14[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm31, %ymm8
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm2[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm2[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm0[4],ymm6[4],ymm0[5],ymm6[5],ymm0[6],ymm6[6],ymm0[7],ymm6[7],ymm0[12],ymm6[12],ymm0[13],ymm6[13],ymm0[14],ymm6[14],ymm0[15],ymm6[15]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm0[4],ymm6[4],ymm0[5],ymm6[5],ymm0[6],ymm6[6],ymm0[7],ymm6[7],ymm0[12],ymm6[12],ymm0[13],ymm6[13],ymm0[14],ymm6[14],ymm0[15],ymm6[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm20, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm2, %xmm11
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm11[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm11 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm12 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm11[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm11
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm12, %zmm11 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm9
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm20
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm9, %ymm14, %ymm20
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm12
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm9 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm9[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm10 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm10, %ymm3, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm21, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm17, %ymm2
-; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm12, %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm29
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm0[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm8
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm0
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm0 {%k1}
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm29, %ymm13
+; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm2, %ymm1
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm23, %ymm9
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm30, %ymm14
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm1
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm18, %ymm15
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm2, %ymm1
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm1[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm2, %ymm6
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm6[0],ymm3[0],ymm6[1],ymm3[1],ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[8],ymm3[8],ymm6[9],ymm3[9],ymm6[10],ymm3[10],ymm6[11],ymm3[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm2[4],ymm5[4],ymm2[5],ymm5[5],ymm2[6],ymm5[6],ymm2[7],ymm5[7],ymm2[12],ymm5[12],ymm2[13],ymm5[13],ymm2[14],ymm5[14],ymm2[15],ymm5[15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqu (%rsp), %ymm4 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm1[4],ymm4[4],ymm1[5],ymm4[5],ymm1[6],ymm4[6],ymm1[7],ymm4[7],ymm1[12],ymm4[12],ymm1[13],ymm4[13],ymm1[14],ymm4[14],ymm1[15],ymm4[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm31, %ymm8
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm10 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm10[0],ymm4[1],ymm10[1],ymm4[2],ymm10[2],ymm4[3],ymm10[3],ymm4[8],ymm10[8],ymm4[9],ymm10[9],ymm4[10],ymm10[10],ymm4[11],ymm10[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm31
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm4, %zmm31 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm7
+; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm7, %ymm4
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm31, %ymm6
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm23, %ymm6
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm22, %ymm31
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm4
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm4[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm4
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm4[2,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm14, %xmm8
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm4, %xmm10
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm8[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm24 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # zmm24 = zmm24[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm24, %zmm0, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm21 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # zmm21 = zmm2[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm24 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm21, %zmm0, %zmm24
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm26[0,1,2,3],zmm18[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm23, %zmm19
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm17, %zmm0, %zmm19
-; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm17 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # zmm17 = zmm27[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm22, %zmm14
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm17, %zmm0, %zmm14
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm4, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm5, %zmm0 {%k2}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm2[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm27, %zmm5
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm28[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm9 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm27, %zmm9
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm4
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm25[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm17 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm17 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm15[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm13, %zmm8
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm6, %ymm13
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsp), %xmm7 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm15 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm15, %ymm23
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm12 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm3
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm3[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm18 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm18 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm13, %zmm8 {%k2}
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm3[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm13
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm6, %zmm31, %zmm13
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm2
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm21 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm29[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm27, %zmm10
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm15 = xmm7[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm15, %zmm8, %zmm31
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm15 = xmm3[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm15, %ymm15
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm3, %xmm3
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm16, %ymm13
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm10, %zmm10
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm13, %zmm10 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm22
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm1
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm29
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm13, %zmm17, %zmm22
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm13, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm15, %xmm12
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm16, %ymm12
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm16
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm12, %zmm16 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm3
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm1, %zmm16, %zmm17
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm1[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm5[0],ymm2[1],ymm5[1],ymm2[2],ymm5[2],ymm2[3],ymm5[3],ymm2[8],ymm5[8],ymm2[9],ymm5[9],ymm2[10],ymm5[10],ymm2[11],ymm5[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm1[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm1[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm20[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20, %zmm20 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm21, %zmm20
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm27[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm27 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm21, %zmm27
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm30, %zmm19
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm21, %zmm19
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm31[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm23, %zmm18
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm6, %zmm21, %zmm18
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm9, %ymm24
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm6 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm9, %ymm11
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm29[0],zero,xmm29[1],zero,xmm29[2],zero,xmm29[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm9, %ymm10
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm9, %ymm16
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm9 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm8, %zmm9 {%k1}
+; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm5 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm13, %zmm5 {%k1}
+; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm12, %zmm8 {%k1}
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm12
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm7, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm21, %ymm9
+; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm3 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, %zmm15
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm7, %zmm15
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm24[0,1,2,3],zmm25[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm25
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm2, %ymm21, %ymm5
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm4 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm14
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm14[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm14, %xmm1
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm28
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm24, %zmm28
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm11[0,1,2,3],zmm26[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm4, %zmm2, %zmm6 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm11
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm4, %zmm7, %zmm11
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm4, %ymm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm2, %xmm3
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm26 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm12
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm20
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm11[0,1,2,3],zmm20[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm20
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm11, %zmm27, %zmm20
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm17, %ymm11, %ymm30
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm18, %ymm11, %ymm16
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm11, %ymm0
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm26, %ymm11, %ymm8
-; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # zmm4 = zmm30[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm23, %zmm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm17, %zmm11
-; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # zmm4 = zmm16[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm17, %zmm2
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm13[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm15, %zmm3
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm17, %zmm3
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm31[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm17, %zmm4
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm25, %xmm26
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm25, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm4
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm21, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm4
+; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm1, %zmm6, %zmm7
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm4[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm1, %ymm25
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm21, %ymm6
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm4, %xmm1
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm1[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm13
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm14, %xmm14
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm2, %xmm2
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm1
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm1[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshuflw $96, (%rsp), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm14[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm14 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm2[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm2 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm4[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm23, %zmm24, %zmm3
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm10[0,1,2,3],zmm22[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm25, %zmm21
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm10, %zmm24, %zmm21
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm16[0,1,2,3],zmm17[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm10, %zmm24, %zmm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm9[0,1,2,3],zmm12[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm26, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm9, %zmm10, %zmm1
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm15[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm29, %zmm9
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm10, %zmm9
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm11[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm30, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm10, %zmm2
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm6[0,1,2,3],zmm7[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm31, %zmm4
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm10, %zmm4
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, 448(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 384(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm11, 576(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 512(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 704(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, 640(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 256(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 448(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 640(%rax)
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-FAST-NEXT: addq $1144, %rsp # imm = 0x478
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, 192(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, 128(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, 320(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 384(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm28, 576(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, 512(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, 704(%rax)
+; AVX512DQ-FAST-NEXT: addq $936, %rsp # imm = 0x3A8
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index f87561e3aaa6..0b578bdd2838 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
@@ -2453,104 +2453,69 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-FAST: # %bb.0:
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm4
; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm0
; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%r10), %ymm2
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm5[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm3[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4],ymm7[5],ymm8[6,7,8,9],ymm7[10],ymm8[11,12],ymm7[13],ymm8[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3],ymm8[4],ymm9[5,6,7,8],ymm8[9],ymm9[10,11],ymm8[12],ymm9[13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,u,u,20,21,24,25,u,u,22,23,22,23]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4],ymm9[5],ymm10[6,7,8,9],ymm9[10],ymm10[11,12],ymm9[13],ymm10[14,15]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <u,5,u,u,u,6,u,u>
-; AVX512F-FAST-NEXT: vpermd %ymm2, %ymm7, %ymm7
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,4,u,u,u,5,u>
-; AVX512F-FAST-NEXT: vpermd %ymm2, %ymm9, %ymm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm9
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm0[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,u,u,18,19,20,21,u,u,20,21]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm10[0,1,2],ymm7[3],ymm10[4,5],ymm7[6],ymm10[7,8,9,10],ymm7[11],ymm10[12,13],ymm7[14],ymm10[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm0[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm7
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm7
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm6[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[12,13,14,15],zero,zero,ymm4[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm4[u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512F-FAST-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm10
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u,u,u,u,u],zero,zero,ymm5[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm5[16,17,u,u,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,14,15],zero,zero,ymm3[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm3[u,u,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpor %ymm11, %ymm12, %ymm13
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm11
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm15
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm15
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[16,17,u,u,u,u]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,ymm5[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[12,13,14,15],zero,zero,ymm3[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm3[u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512F-FAST-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u],zero,zero,ymm4[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[16,17,u,u,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpor %ymm7, %ymm10, %ymm7
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm12, %zmm15
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm15
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm1[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[16,17,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-FAST-NEXT: vpor %ymm8, %ymm13, %ymm8
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm13
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm14
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm1, %xmm19
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpbroadcastd 8(%r10), %ymm8
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandnq %ymm8, %ymm16, %ymm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[16,17,u,u]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm8
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm8
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512F-FAST-NEXT: vprold $16, %xmm10, %xmm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm9[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0,1],xmm1[2],xmm15[3,4],xmm1[5],xmm15[6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm1 = xmm11[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm12[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm11[0],xmm1[1],xmm11[2,3],xmm1[4],xmm11[5,6],xmm1[7]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm19, %xmm10
-; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm9, %xmm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512F-FAST-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm12
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm13
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm16
+; AVX512F-FAST-NEXT: vpbroadcastd 8(%r10), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm6, %ymm7, %ymm6
+; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[16,17,u,u]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,1,3]
+; AVX512F-FAST-NEXT: vprold $16, %xmm9, %xmm15
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm8[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm15[2],xmm0[3,4],xmm15[5],xmm0[6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm0
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm18, %xmm9
+; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm8, %xmm8
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm0[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
@@ -2560,29 +2525,60 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-FAST-NEXT: vpbroadcastd 4(%r10), %ymm10
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm9
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,14,15,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1],ymm0[2],ymm8[3,4],ymm0[5],ymm8[6,7,8,9],ymm0[10],ymm8[11,12],ymm0[13],ymm8[14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2,3],ymm8[4],ymm10[5,6,7,8],ymm8[9],ymm10[10,11],ymm8[12],ymm10[13,14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,2,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0,1],ymm8[2],ymm10[3,4],ymm8[5],ymm10[6,7,8,9],ymm8[10],ymm10[11,12],ymm8[13],ymm10[14,15]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,u,u,20,21,24,25,u,u,22,23,22,23]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm17[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,u,u,18,19,20,21,u,u,20,21]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1,2],ymm0[3],ymm10[4,5],ymm0[6],ymm10[7,8,9,10],ymm0[11],ymm10[12,13],ymm0[14],ymm10[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm10
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm17[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512F-FAST-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-FAST-NEXT: vpermd %zmm7, %zmm10, %zmm10
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm10
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,14,15,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u]
; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5],ymm0[6],ymm3[7,8,9,10],ymm0[11],ymm3[12,13],ymm0[14],ymm3[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <6,u,u,u,7,u,u,7>
-; AVX512F-FAST-NEXT: vpermd %ymm2, %ymm0, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm2, 192(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512F-FAST-NEXT: vpermd %ymm7, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa %ymm1, 192(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, 128(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 64(%rax)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
@@ -5336,330 +5332,321 @@ define void @store_i16_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-FAST-LABEL: store_i16_stride7_vf32:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: subq $616, %rsp # imm = 0x268
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm11
+; AVX512F-FAST-NEXT: subq $712, %rsp # imm = 0x2C8
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm11, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm10
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm2
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm2, %ymm2
; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm29
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm4, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm28
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm5
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm5, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm2
-; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm23
-; AVX512F-FAST-NEXT: vprold $16, %ymm5, %ymm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7,8,9,10],ymm4[11],ymm2[12,13],ymm4[14],ymm2[15]
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,2,3,3,10,9,11,10]
-; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm7
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm1
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm4, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm16
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm14
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vpor %ymm2, %ymm4, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm3, %ymm3
+; AVX512F-FAST-NEXT: vpor %ymm1, %ymm3, %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm9
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm3
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm10
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm10, %ymm4
+; AVX512F-FAST-NEXT: vporq %ymm3, %ymm4, %ymm19
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm7
+; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm12
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm12, %ymm2
+; AVX512F-FAST-NEXT: vpor %ymm4, %ymm2, %ymm2
; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm15, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm7
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm7, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm14, %ymm2
; AVX512F-FAST-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512F-FAST-NEXT: vporq %ymm0, %ymm2, %ymm17
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm12
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa %xmm1, %xmm8
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm1
-; AVX512F-FAST-NEXT: vpermi2q %zmm0, %zmm1, %zmm3
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm4
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,0,0,1,8,9,9,11]
-; AVX512F-FAST-NEXT: vpermi2q %zmm4, %zmm3, %zmm9
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm0, %zmm1
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = <0,0,1,1,12,13,u,15>
-; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm0, %zmm23
-; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm15
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm15, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm2
+; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm17
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,1,1,3,8,8,9,9]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm2
+; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm5
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm4
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,0,0,1,8,9,9,11]
+; AVX512F-FAST-NEXT: vpermi2q %zmm5, %zmm4, %zmm11
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm0, %zmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = <0,0,1,1,12,13,u,15>
+; AVX512F-FAST-NEXT: vpermi2q %zmm2, %zmm1, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31>
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm5, %ymm2
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm1
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm9, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[3,3,3,3,7,7,7,7]
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm6
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [2,2,2,3,8,8,8,9]
-; AVX512F-FAST-NEXT: vpermi2q %zmm0, %zmm2, %zmm26
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm6, %ymm2
-; AVX512F-FAST-NEXT: vpbroadcastd 8(%rax), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandn %ymm3, %ymm4, %ymm3
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm27
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm11[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm10, %ymm19
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [2,2,2,3,8,8,8,9]
+; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vprold $16, %ymm9, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm10[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21>
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm9, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0,1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7,8,9,10],ymm3[11],ymm1[12,13],ymm3[14],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm11
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [2,2,3,3,10,9,11,10]
+; AVX512F-FAST-NEXT: vpermi2q %zmm2, %zmm3, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm3
+; AVX512F-FAST-NEXT: vpbroadcastd 8(%rax), %ymm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm4, %ymm7, %ymm4
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm28
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27>
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm11, %ymm9
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm30
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23>
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm29, %ymm10
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm10, %ymm3
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm31
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm4
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm28, %ymm29
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm16, %ymm5
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm2
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm4
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm18[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm12[3],ymm2[4,5],ymm12[6],ymm2[7,8,9,10],ymm12[11],ymm2[12,13],ymm12[14],ymm2[15]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm7[0,1],ymm3[2],ymm7[3,4],ymm3[5],ymm7[6,7,8,9],ymm3[10],ymm7[11,12],ymm3[13],ymm7[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm30
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm15, %ymm5
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = <0,1,u,3,10,10,11,11>
-; AVX512F-FAST-NEXT: vpermi2q %zmm2, %zmm17, %zmm25
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,4,5,4,5,5,7]
-; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm12
-; AVX512F-FAST-NEXT: vpermd %ymm12, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandn %ymm2, %ymm13, %ymm2
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm28
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm15, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm15, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vprold $16, %ymm4, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm4, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512F-FAST-NEXT: vprold $16, %xmm2, %xmm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2],xmm3[3,4],xmm0[5],xmm3[6,7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm8
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm13
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm5
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm14, %xmm0
+; AVX512F-FAST-NEXT: vpermi2q %zmm5, %zmm17, %zmm25
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,4,u,u,u,5,u>
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm9
+; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm5, %ymm5
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm5, %ymm7, %ymm5
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm9, %ymm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm17
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6,7,8],ymm2[9],ymm7[10,11],ymm2[12],ymm7[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm31
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm2[3],ymm6[4,5],ymm2[6],ymm6[7,8,9,10],ymm2[11],ymm6[12,13],ymm2[14],ymm6[15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm12[2],ymm2[3,4],ymm12[5],ymm2[6,7,8,9],ymm12[10],ymm2[11,12],ymm12[13],ymm2[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm13, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm13[0],ymm2[1],ymm13[2,3],ymm2[4],ymm13[5,6,7,8],ymm2[9],ymm13[10,11],ymm2[12],ymm13[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm24
+; AVX512F-FAST-NEXT: vprold $16, %ymm15, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm14[0,1],ymm2[2],ymm14[3,4],ymm2[5],ymm14[6,7,8,9],ymm2[10],ymm14[11,12],ymm2[13],ymm14[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm15, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7,8,9],ymm0[10],ymm2[11,12],ymm0[13],ymm2[14,15]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm14
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9>
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm1[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm15[0],xmm0[1],xmm15[2,3],xmm0[4],xmm15[5,6],xmm0[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3]
-; AVX512F-FAST-NEXT: vprold $16, %xmm13, %xmm13
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0,1],xmm13[2],xmm8[3,4],xmm13[5],xmm8[6,7]
+; AVX512F-FAST-NEXT: vmovdqa64 (%rax), %zmm27
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm29 = <u,5,u,u,u,6,u,u,30,u,u,u,31,u,u,31>
+; AVX512F-FAST-NEXT: vpermi2d %zmm27, %zmm9, %zmm29
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512F-FAST-NEXT: vprold $16, %xmm10, %xmm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm9[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0,1],xmm2[2],xmm8[3,4],xmm2[5],xmm8[6,7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm2, %xmm6
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm2
+; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = <u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9>
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm5, %xmm3
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm2[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm8[0],xmm3[1],xmm8[2,3],xmm3[4],xmm8[5,6],xmm3[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm13, %xmm3
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm1, %xmm4
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm3[1],xmm0[2,3],xmm3[4],xmm0[5,6],xmm3[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm9, %ymm3
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm19[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm11[0],ymm3[1],ymm11[2,3],ymm3[4],ymm11[5,6,7,8],ymm3[9],ymm11[10,11],ymm3[12],ymm11[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm1[4],xmm14[4],xmm1[5],xmm14[5],xmm1[6],xmm14[6],xmm1[7],xmm14[7]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1],xmm1[2],xmm14[2],xmm1[3],xmm14[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm13, %xmm13
-; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm1, %xmm9
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm29[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1,2],ymm14[3],ymm10[4,5],ymm14[6],ymm10[7,8,9,10],ymm14[11],ymm10[12,13],ymm14[14],ymm10[15]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm15, %xmm7
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm8
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [4,5,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [6,7,3,3,7,7,6,7]
-; AVX512F-FAST-NEXT: vpermd %ymm12, %ymm15, %ymm29
-; AVX512F-FAST-NEXT: vpermd %ymm12, %ymm14, %ymm12
-; AVX512F-FAST-NEXT: vpermd %ymm6, %ymm14, %ymm14
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [0,1,4,5,4,5,5,7]
-; AVX512F-FAST-NEXT: vpermd %ymm6, %ymm0, %ymm19
-; AVX512F-FAST-NEXT: vpermd %ymm6, %ymm15, %ymm0
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm15 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm0 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3],xmm4[4],xmm0[5,6],xmm4[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm23
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm0, %ymm4
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm14[0],ymm4[1],ymm14[2,3],ymm4[4],ymm14[5,6,7,8],ymm4[9],ymm14[10,11],ymm4[12],ymm14[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm26
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm2, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm15[0,1,2],ymm5[3],ymm15[4,5],ymm5[6],ymm15[7,8,9,10],ymm5[11],ymm15[12,13],ymm5[14],ymm15[15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm16
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm8, %xmm8
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm9, %xmm9
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm15[0,1],ymm10[2],ymm15[3,4],ymm10[5],ymm15[6,7,8,9],ymm10[10],ymm15[11,12],ymm10[13],ymm15[14,15]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm13 = ymm15[0,1],ymm13[2],ymm15[3,4],ymm13[5],ymm15[6,7,8,9],ymm13[10],ymm15[11,12],ymm13[13],ymm15[14,15]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm2[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm15[0,1],ymm7[2],ymm15[3,4],ymm7[5],ymm15[6,7,8,9],ymm7[10],ymm15[11,12],ymm7[13],ymm15[14,15]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm1[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0],ymm15[1],ymm12[2,3],ymm15[4],ymm12[5,6,7,8],ymm15[9],ymm12[10,11],ymm15[12],ymm12[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <6,u,u,u,7,u,u,7>
+; AVX512F-FAST-NEXT: vpermd %ymm11, %ymm15, %ymm15
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm27, %zmm11
+; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm27 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,1,3,2]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $246, (%rsp), %ymm2 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm2 = mem[2,1,3,3]
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
; AVX512F-FAST-NEXT: # ymm1 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm2 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm6 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm24[2,1,3,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm27
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm27
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm19, %zmm4
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm15, %zmm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm31, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm0
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
-; AVX512F-FAST-NEXT: vpermq $234, (%rsp), %ymm4 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm4 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm5 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm14 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm15 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm22[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm20[0,0,1,1]
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm21[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,1,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm17[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm16[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,1,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm24[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm5 = mem[2,1,3,2]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm22[2,2,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm24, %zmm6
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm28
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm28
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm3, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm17
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm17
+; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm3 = mem[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm6 = mem[0,0,1,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm19[0,0,1,1]
+; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm21, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm28
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm28
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm2
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm3, %zmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm1
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm18[0,1,1,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm26[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm27 = ymm20[0,1,1,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm16[2,1,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,2,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm31, %zmm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm29
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm19, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm24, %zmm2
+; AVX512F-FAST-NEXT: vpbroadcastd 36(%rax), %ymm0
+; AVX512F-FAST-NEXT: vpbroadcastd 40(%rax), %ymm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm8, %zmm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm22, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm25, %zmm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm30, %zmm4
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm1, %zmm4
+; AVX512F-FAST-NEXT: vpbroadcastd (%rax), %ymm1
+; AVX512F-FAST-NEXT: vpbroadcastd 4(%rax), %ymm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm12, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm23, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm4
-; AVX512F-FAST-NEXT: vpbroadcastd 36(%rax), %ymm1
-; AVX512F-FAST-NEXT: vpbroadcastd 40(%rax), %ymm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm1
+; AVX512F-FAST-NEXT: vpbroadcastd 32(%rax), %ymm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm15, %zmm2
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm10, %zmm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm4
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = <u,u,4,u,u,u,5,u,u,13,u,u,u,14,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm11, %zmm3, %zmm3
; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm1
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm7, %zmm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm13, %zmm5
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm31, %zmm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm3, %zmm6
-; AVX512F-FAST-NEXT: vpbroadcastd (%rax), %ymm3
-; AVX512F-FAST-NEXT: vpbroadcastd 4(%rax), %ymm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm3
-; AVX512F-FAST-NEXT: vpbroadcastd 32(%rax), %ymm4
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm28, 320(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 128(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm27, 64(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 384(%rax)
-; AVX512F-FAST-NEXT: addq $616, %rsp # imm = 0x268
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, 320(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm28, 64(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, (%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 256(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm29, 384(%rax)
+; AVX512F-FAST-NEXT: addq $712, %rsp # imm = 0x2C8
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
@@ -11431,742 +11418,724 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-FAST-LABEL: store_i16_stride7_vf64:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: subq $2392, %rsp # imm = 0x958
-; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %ymm12
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %ymm0
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %ymm10
+; AVX512F-FAST-NEXT: subq $2200, %rsp # imm = 0x898
+; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %ymm15
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa %ymm2, %ymm5
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm7
-; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm19
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm12, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa %ymm2, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm4, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm18
-; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm11
-; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm17
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm3
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm14
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512F-FAST-NEXT: vprold $16, %ymm3, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm15
-; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [2,1,3,2,10,10,10,11]
-; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm3
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm2, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm22
-; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm3, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm4, %ymm9
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm23
-; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm26
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm3, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm15, %ymm3
+; AVX512F-FAST-NEXT: vporq %ymm0, %ymm3, %ymm17
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm6, %ymm5
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm6, %ymm23
+; AVX512F-FAST-NEXT: vporq %ymm4, %ymm5, %ymm16
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa %ymm5, %ymm6
+; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm5
+; AVX512F-FAST-NEXT: vporq %ymm4, %ymm5, %ymm24
; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %ymm2
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa %ymm7, %ymm4
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm5
+; AVX512F-FAST-NEXT: vpor %ymm4, %ymm5, %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %ymm7
-; AVX512F-FAST-NEXT: vmovdqa %ymm11, %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm7, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm31
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm30
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %ymm14
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm14, %ymm5
+; AVX512F-FAST-NEXT: vpor %ymm4, %ymm5, %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa %ymm8, %ymm7
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm2
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm9, %ymm11
-; AVX512F-FAST-NEXT: vporq %ymm1, %ymm2, %ymm27
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa %ymm6, %ymm10
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm5
+; AVX512F-FAST-NEXT: vporq %ymm4, %ymm5, %ymm31
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa %ymm5, %ymm9
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm5
+; AVX512F-FAST-NEXT: vpor %ymm4, %ymm5, %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512F-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm5
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm1, %ymm5
+; AVX512F-FAST-NEXT: vpor %ymm4, %ymm5, %ymm1
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm13
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm13, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm3, %ymm2
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm12, %ymm4
+; AVX512F-FAST-NEXT: vpor %ymm2, %ymm4, %ymm1
; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm2, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm5
-; AVX512F-FAST-NEXT: vpor %ymm4, %ymm5, %ymm4
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm14, %ymm28
-; AVX512F-FAST-NEXT: vmovdqa %ymm7, %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm15, %ymm5
-; AVX512F-FAST-NEXT: vporq %ymm5, %ymm4, %ymm29
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm12, %ymm4
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29>
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm12, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa %ymm12, %ymm7
-; AVX512F-FAST-NEXT: vmovdqa %ymm5, %ymm8
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27>
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm4
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa %ymm5, %ymm12
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm11, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm7
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm7, %ymm0
+; AVX512F-FAST-NEXT: vpor %ymm2, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm6
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm9
+; AVX512F-FAST-NEXT: vporq %ymm9, %ymm0, %ymm22
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7,8,9],ymm9[10],ymm0[11,12],ymm9[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vprold $16, %ymm5, %ymm9
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7,8,9],ymm9[10],ymm8[11,12],ymm9[13],ymm8[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [2,1,3,2,10,10,10,11]
+; AVX512F-FAST-NEXT: vpermi2q %zmm0, %zmm8, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u>
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm29
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm23[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2,3],ymm0[4],ymm8[5,6,7,8],ymm0[9],ymm8[10,11],ymm0[12],ymm8[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29>
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm23[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1,2],ymm0[3],ymm8[4,5],ymm0[6],ymm8[7,8,9,10],ymm0[11],ymm8[12,13],ymm0[14],ymm8[15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27>
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm4
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm4, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm15[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm8[2],ymm0[3,4],ymm8[5],ymm0[6,7,8,9],ymm8[10],ymm0[11,12],ymm8[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm4, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2,3],ymm0[4],ymm8[5,6,7,8],ymm0[9],ymm8[10,11],ymm0[12],ymm8[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm15[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1],ymm0[2],ymm8[3,4],ymm0[5],ymm8[6,7,8,9],ymm0[10],ymm8[11,12],ymm0[13],ymm8[14,15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23>
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm7, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm5, %ymm16
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm19, %zmm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm17, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23>
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm8
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm23[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7,8,9],ymm9[10],ymm8[11,12],ymm9[13],ymm8[14,15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm17, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm16, %zmm8
+; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <5,u,u,u,6,u,u,6>
-; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %ymm5
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %ymm9
+; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm0, %ymm0
; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,4,u,u,u,5,u,u>
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm0, %ymm0
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[14,15],zero,zero,ymm5[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[16,17],zero,zero,ymm5[u,u],zero,zero
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $248, %ymm19, %ymm4, %ymm5
-; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %ymm7
-; AVX512F-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm7, %ymm10
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm10
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm4, %ymm4
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm0
-; AVX512F-FAST-NEXT: vprold $16, %ymm7, %ymm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm10[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm9[14,15],zero,zero,ymm9[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm9[16,17],zero,zero,ymm9[u,u],zero,zero
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $248, %ymm16, %ymm8, %ymm9
+; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %ymm1
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm1, %ymm15
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm15
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm8, %ymm8
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm0
+; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm8
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm8
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm15[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [0,1,4,5,4,5,5,7]
-; AVX512F-FAST-NEXT: vmovdqa 96(%rax), %ymm4
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpermd %ymm4, %ymm20, %ymm0
-; AVX512F-FAST-NEXT: vpandnq %ymm0, %ymm19, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm4, %ymm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 64(%rax), %zmm23
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,5,u,u,u,6,u,u,30,u,u,u,31,u,u,31>
+; AVX512F-FAST-NEXT: vmovdqa 96(%rax), %ymm0
+; AVX512F-FAST-NEXT: vpermi2d %zmm23, %zmm0, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [0,1,4,5,4,5,5,7]
+; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm8
+; AVX512F-FAST-NEXT: vpandnq %ymm8, %ymm16, %ymm8
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0],ymm0[1],ymm4[2,3],ymm0[4],ymm4[5,6,7,8],ymm0[9],ymm4[10,11],ymm0[12],ymm4[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm2, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5],ymm0[6],ymm4[7,8,9,10],ymm0[11],ymm4[12,13],ymm0[14],ymm4[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm15, %ymm25
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7,8,9],ymm4[10],ymm0[11,12],ymm4[13],ymm0[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm13, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm13, %ymm24
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0],ymm0[1],ymm4[2,3],ymm0[4],ymm4[5,6,7,8],ymm0[9],ymm4[10,11],ymm0[12],ymm4[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
; AVX512F-FAST-NEXT: vpbroadcastd 72(%rax), %ymm0
-; AVX512F-FAST-NEXT: vpandn %ymm0, %ymm8, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa 64(%rax), %ymm0
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm4, %zmm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm4
-; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %xmm6
-; AVX512F-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %xmm7
-; AVX512F-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm10
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [0,0,1,1,12,13,14,15]
-; AVX512F-FAST-NEXT: vpermt2q %zmm4, %zmm18, %zmm10
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm21, %zmm10, %zmm5
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm22, %ymm15
-; AVX512F-FAST-NEXT: vprold $16, %ymm22, %ymm4
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm23[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm4[2],ymm10[3,4],ymm4[5],ymm10[6,7,8,9],ymm4[10],ymm10[11,12],ymm4[13],ymm10[14,15]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21>
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm15, %ymm13
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm22, %ymm30
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm23[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm23, %ymm26
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1,2],ymm15[3],ymm13[4,5],ymm15[6],ymm13[7,8,9,10],ymm15[11],ymm13[12,13],ymm15[14],ymm13[15]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = [2,2,3,3,10,9,11,10]
-; AVX512F-FAST-NEXT: vpermt2q %zmm10, %zmm22, %zmm13
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [4,5,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm10
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm15
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm15, %zmm5
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm13, %zmm23, %zmm5
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm16, %ymm4
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm7, %ymm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm31[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm7, %ymm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm31[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm31, %ymm16
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm13[0],ymm10[1],ymm13[2,3],ymm10[4],ymm13[5,6,7,8],ymm10[9],ymm13[10,11],ymm10[12],ymm13[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm5[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm31[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm13[0,1],ymm10[2],ymm13[3,4],ymm10[5],ymm13[6,7,8,9],ymm10[10],ymm13[11,12],ymm10[13],ymm13[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm10
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm31[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpbroadcastd 8(%rax), %ymm10
-; AVX512F-FAST-NEXT: vpandn %ymm10, %ymm8, %ymm13
-; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm10
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm10, %ymm15
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm13, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm27
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm8
+; AVX512F-FAST-NEXT: vpandn %ymm0, %ymm1, %ymm8
+; AVX512F-FAST-NEXT: vmovdqa 64(%rax), %ymm10
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm10, %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm27
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm15
+; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %xmm9
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm2[4],xmm9[5],xmm2[5],xmm9[6],xmm2[6],xmm9[7],xmm2[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm28
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm24 = [0,0,1,1,12,13,14,15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm15, %zmm24, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm26, %zmm0, %zmm27
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpbroadcastd 8(%rax), %ymm0
+; AVX512F-FAST-NEXT: vpandn %ymm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm15
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm8
; AVX512F-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm13
-; AVX512F-FAST-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm15, %xmm15
-; AVX512F-FAST-NEXT: vpermt2q %zmm27, %zmm18, %zmm15
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm21, %zmm15, %zmm6
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512F-FAST-NEXT: vprold $16, %ymm12, %ymm15
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm6[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4],ymm15[5],ymm13[6,7,8,9],ymm15[10],ymm13[11,12],ymm15[13],ymm13[14,15]
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm12, %ymm15
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm12, %ymm27
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm12 = ymm15[0,1,2],ymm12[3],ymm15[4,5],ymm12[6],ymm15[7,8,9,10],ymm12[11],ymm15[12,13],ymm12[14],ymm15[15]
-; AVX512F-FAST-NEXT: vpermt2q %zmm13, %zmm22, %zmm12
-; AVX512F-FAST-NEXT: vpermd %ymm10, %ymm17, %ymm13
-; AVX512F-FAST-NEXT: vpermd %ymm10, %ymm20, %ymm15
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm12, %zmm23, %zmm6
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu (%rsp), %ymm15 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm15, %ymm12
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm8[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm9
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm15, %ymm23
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm8[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm12[0],ymm9[1],ymm12[2,3],ymm9[4],ymm12[5,6,7,8],ymm9[9],ymm12[10,11],ymm9[12],ymm12[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm18[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm12[0,1],ymm9[2],ymm12[3,4],ymm9[5],ymm12[6,7,8,9],ymm9[10],ymm12[11,12],ymm9[13],ymm12[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm15, %ymm8
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm18[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7,8,9],ymm9[10],ymm8[11,12],ymm9[13],ymm8[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm24, %ymm6
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm28[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,1,u,3,10,10,11,11>
-; AVX512F-FAST-NEXT: vpermi2q %zmm1, %zmm29, %zmm2
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm20, %ymm1
-; AVX512F-FAST-NEXT: vpandnq %ymm1, %ymm19, %ymm1
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm9, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm4[4],xmm8[5],xmm4[5],xmm8[6],xmm4[6],xmm8[7],xmm4[7]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm15, %xmm15
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm24, %zmm15
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm26, %zmm15, %zmm1
; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm31[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm15[0,1],ymm0[2],ymm15[3,4],ymm0[5],ymm15[6,7,8,9],ymm0[10],ymm15[11,12],ymm0[13],ymm15[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm18, %ymm4
+; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm11, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm7[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4],ymm15[5],ymm0[6,7,8,9],ymm15[10],ymm0[11,12],ymm15[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21>
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm5, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8,9,10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <0,1,u,3,10,10,11,11>
+; AVX512F-FAST-NEXT: vpermi2q %zmm0, %zmm22, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm0
+; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm2
+; AVX512F-FAST-NEXT: vpandnq %ymm2, %ymm16, %ymm2
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm29, %ymm5
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm11, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm7[2,2,2,2,6,6,6,6]
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm20, %ymm6
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm11, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31>
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm30, %ymm1
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm26[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm21, %ymm3
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm13, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[2,2,2,2,6,6,6,6]
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [2,2,2,3,8,8,8,9]
-; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %xmm2
-; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %xmm3
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm4, %xmm5
-; AVX512F-FAST-NEXT: vpermt2q %zmm5, %zmm6, %zmm1
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [6,7,3,3,7,7,6,7]
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm14, %ymm0
-; AVX512F-FAST-NEXT: vpbroadcastd 96(%rax), %ymm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm15
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 (%rax), %zmm21
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,21,u,u,u,22,u,u,14,u,u,u,15,u,u,15>
+; AVX512F-FAST-NEXT: vpermi2d %zmm0, %zmm21, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm30, %ymm2
+; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm30, %ymm4
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm5[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa %ymm5, %ymm11
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vmovdqa %ymm2, %ymm7
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512F-FAST-NEXT: vprold $16, %ymm2, %ymm0
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm5[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa %ymm2, %ymm3
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [2,2,3,3,10,9,11,10]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm12, %zmm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm23, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = <u,u,4,u,u,u,5,u,u,13,u,u,u,14,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm11, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31>
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [2,2,2,3,8,8,8,9]
+; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %xmm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm2
+; AVX512F-FAST-NEXT: vmovdqa %xmm3, %xmm14
+; AVX512F-FAST-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [6,7,3,3,7,7,6,7]
+; AVX512F-FAST-NEXT: vpermd %ymm10, %ymm2, %ymm3
+; AVX512F-FAST-NEXT: vpbroadcastd 96(%rax), %ymm4
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm1
-; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm5
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
-; AVX512F-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm5, %xmm31
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm29
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm4
+; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm10
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm4[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm10[2],xmm11[3,4],xmm10[5],xmm11[6,7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm23
; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = <u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9>
-; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm5
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512F-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm10
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = <u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9>
+; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm4
+; AVX512F-FAST-NEXT: vmovdqa %xmm15, %xmm8
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm10[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0],xmm4[1],xmm11[2,3],xmm4[4],xmm11[5,6],xmm4[7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm0[4],xmm10[5],xmm0[5],xmm10[6],xmm0[6],xmm10[7],xmm0[7]
; AVX512F-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} xmm20 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm1
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,1,1,3,8,8,9,9]
-; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm7, %xmm5
+; AVX512F-FAST-NEXT: vmovdqa %xmm15, %xmm7
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm3, %zmm5
; AVX512F-FAST-NEXT: vpbroadcastd 100(%rax), %ymm0
-; AVX512F-FAST-NEXT: vpbroadcastd 104(%rax), %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vpbroadcastd 104(%rax), %ymm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm30
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512F-FAST-NEXT: vmovdqa %xmm13, %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm1, %xmm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm13, %xmm16
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512F-FAST-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512F-FAST-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm2
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [0,0,0,1,8,9,9,11]
-; AVX512F-FAST-NEXT: vpermt2q %zmm2, %zmm19, %zmm1
-; AVX512F-FAST-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512F-FAST-NEXT: vpbroadcastd 68(%rax), %ymm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm4
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm4
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm5
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm6
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm1, %xmm17
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm5
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; AVX512F-FAST-NEXT: vmovdqa %xmm8, %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm5[0],xmm0[1],xmm5[2,3],xmm0[4],xmm5[5,6],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %xmm5
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm15, %xmm17
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
+; AVX512F-FAST-NEXT: vprold $16, %xmm5, %xmm5
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm5[2],xmm0[3,4],xmm5[5],xmm0[6,7]
; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm5 # 16-byte Folded Reload
-; AVX512F-FAST-NEXT: # xmm5 = xmm4[0],mem[0],xmm4[1],mem[1],xmm4[2],mem[2],xmm4[3],mem[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm6
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vpermt2q %zmm6, %zmm19, %zmm5
-; AVX512F-FAST-NEXT: vpbroadcastd (%rax), %ymm3
-; AVX512F-FAST-NEXT: vpbroadcastd 4(%rax), %ymm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm2, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm28, %xmm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
+; AVX512F-FAST-NEXT: vmovdqa %xmm7, %xmm13
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm0, %xmm5
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,0,0,1,8,9,9,11]
+; AVX512F-FAST-NEXT: vpermt2q %zmm5, %zmm6, %zmm0
+; AVX512F-FAST-NEXT: vpbroadcastd 64(%rax), %ymm5
+; AVX512F-FAST-NEXT: vpbroadcastd 68(%rax), %ymm7
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm3
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm5
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm3, %xmm16
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm5[0],xmm0[1],xmm5[2,3],xmm0[4],xmm5[5,6],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm28
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
+; AVX512F-FAST-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm9[2],xmm0[3,4],xmm9[5],xmm0[6,7]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm27
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm9
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm13, %xmm29
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa %xmm14, %xmm2
+; AVX512F-FAST-NEXT: vpermt2q %zmm9, %zmm6, %zmm0
+; AVX512F-FAST-NEXT: vpbroadcastd (%rax), %ymm6
+; AVX512F-FAST-NEXT: vpbroadcastd 4(%rax), %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm26
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm14
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[2,3],ymm0[4],ymm6[5,6,7,8],ymm0[9],ymm6[10,11],ymm0[12],ymm6[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-FAST-NEXT: vprold $16, %ymm3, %ymm0
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm9[1,2,2,3,5,6,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm9[0,0,2,1,4,4,6,5]
+; AVX512F-FAST-NEXT: vmovdqa %ymm9, %ymm10
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm8[3],ymm6[4,5],ymm8[6],ymm6[7,8,9,10],ymm8[11],ymm6[12,13],ymm8[14],ymm6[15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm12, %zmm6
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1],ymm0[2],ymm8[3,4],ymm0[5],ymm8[6,7,8,9],ymm0[10],ymm8[11,12],ymm0[13],ymm8[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm24
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm8[2],ymm0[3,4],ymm8[5],ymm0[6,7,8,9],ymm8[10],ymm0[11,12],ymm8[13],ymm0[14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm21, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,u,4,u,u,u,5,u,u,13,u,u,u,14,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm8, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm6, %zmm1, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm25
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
-; AVX512F-FAST-NEXT: vprold $16, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm9[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm15[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm27, %ymm2
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vpshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm5 = mem[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm12 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm6
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm5, %xmm8
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm11, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm7, %xmm12
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm1[3],ymm4[4,5],ymm1[6],ymm4[7,8,9,10],ymm1[11],ymm4[12,13],ymm1[14],ymm4[15]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm10[3,3,3,3,7,7,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm11
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,2,2,3,8,8,8,9]
-; AVX512F-FAST-NEXT: vpermt2q %zmm8, %zmm0, %zmm12
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm11, %xmm0
+; AVX512F-FAST-NEXT: vpermt2q %zmm11, %zmm0, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [6,7,3,3,7,7,6,7]
+; AVX512F-FAST-NEXT: vpermd %ymm19, %ymm0, %ymm2
+; AVX512F-FAST-NEXT: vpbroadcastd 32(%rax), %ymm11
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm22
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm22
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm7, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm26
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm1, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm27
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm11, %xmm4
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm23, %ymm7
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512F-FAST-NEXT: vpshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm15 = mem[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm15[0,1,2],ymm11[3],ymm15[4,5],ymm11[6],ymm15[7,8,9,10],ymm11[11],ymm15[12,13],ymm11[14],ymm15[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpermd %ymm7, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [4,5,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpermd %ymm7, %ymm4, %ymm7
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm4, %ymm4
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpermd %ymm9, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpermd %ymm10, %ymm14, %ymm7
-; AVX512F-FAST-NEXT: vpbroadcastd 32(%rax), %ymm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm24
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm24
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm31, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm19
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm9
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm10, %xmm3
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm15, %xmm8
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm11
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm3
; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm6, %xmm6
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm5[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm4
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,1,1,3,8,8,9,9]
-; AVX512F-FAST-NEXT: vpermt2q %zmm6, %zmm2, %zmm14
-; AVX512F-FAST-NEXT: vpbroadcastd 36(%rax), %ymm5
-; AVX512F-FAST-NEXT: vpbroadcastd 40(%rax), %ymm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm5
-; AVX512F-FAST-NEXT: vprold $16, %xmm11, %xmm6
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm14[0,1],xmm6[2],xmm14[3,4],xmm6[5],xmm14[6,7]
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm13
-; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm13, %xmm7
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm25, %xmm2
-; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm29, %xmm3
-; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm3, %xmm3
-; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm9, %xmm3
-; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm1[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm0[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm10[0],xmm9[1],xmm10[2,3],xmm9[4],xmm10[5,6],xmm9[7]
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm4
+; AVX512F-FAST-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512F-FAST-NEXT: vpbroadcastd 40(%rax), %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm10
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm10
+; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm7[1,1,2,3]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0,1],xmm1[2],xmm4[3,4],xmm1[5],xmm4[6,7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm17, %xmm15
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm15, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm2
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm23, %xmm3
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm3
; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm10 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm29 = mem[2,1,3,3]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm4 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm3 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
-; AVX512F-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm0 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpermpd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm0 = mem[2,1,3,3]
-; AVX512F-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm0 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm6[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm4[0],xmm0[1],xmm4[2,3],xmm0[4],xmm4[5,6],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm23 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm4 = mem[2,1,3,3]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm3 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
; AVX512F-FAST-NEXT: # ymm0 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm30, %xmm0
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7]
+; AVX512F-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufhw {{.*#+}} ymm9 = ymm5[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm31, %xmm5
+; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm5[0,2,3,3,4,5,6,7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm20 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm16 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm15 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm13 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm17, %xmm0
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm5
+; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm9 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm8 = mem[0,2,2,3]
; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
; AVX512F-FAST-NEXT: # ymm6 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm0 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm17 = mem[2,1,3,2]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm18 = mem[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm26[0,1,1,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm28[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm19[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm27[0,1,1,3]
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm27 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm22[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm22 = mem[0,0,2,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm10, %zmm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm29 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm10, %zmm29, %zmm3
-; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm4
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm12
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm3, %ymm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm4, %ymm3
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm12[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, (%rsp), %zmm3, %zmm19 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm1, %zmm10, %zmm19
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm3 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm10 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm10, %zmm26, %zmm7
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm7 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm7 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm5 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vmovups %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpermpd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm5 = mem[2,1,3,3]
+; AVX512F-FAST-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm5 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vmovups %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm5 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm31 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm30 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm30 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm29 = mem[2,1,3,2]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm28 = mem[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm21[0,1,1,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm8[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm8 = mem[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm12[0,1,1,3]
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm12 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm11[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm27[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm27 = ymm26[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm25[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm24[2,1,3,2]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm20[2,2,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm23, %zmm4
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm3, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm20
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm20, %ymm15
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm4 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm23, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm13 # 32-byte Folded Reload
; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm11, %zmm26, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm2 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm11
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm10, %zmm13, %zmm11
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm8, %zmm13, %zmm0
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm11 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm0 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm4, %zmm6
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm4, %zmm8
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm14 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm1 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm31, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm28, %zmm9
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm10, %zmm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm30, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm25, %zmm13
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm10, %zmm13
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm9, %zmm6, %zmm27
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm13, %zmm6, %zmm28
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm13, %zmm23, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm6 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm9 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm29, %zmm9
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm6 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm10 = mem[0,1,1,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
-; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm10 = mem[2,1,3,3]
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm13 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm10, %zmm10
-; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm13 = mem[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm25 = mem[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm15 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
-; AVX512F-FAST-NEXT: # xmm4 = mem[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm16 = mem[0,2,2,3]
-; AVX512F-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm17 = mem[0,1,1,3]
-; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm18 = mem[2,1,3,3]
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm20 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm21 = mem[0,0,2,1]
-; AVX512F-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm22 = mem[0,0,1,3]
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm23 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512F-FAST-NEXT: # xmm12 = mem[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm29, %zmm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm16, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm18, %zmm16
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm29, %zmm16
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm20 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm0, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm4, %ymm0
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm15[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm13 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm19, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm17, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm7, %zmm4
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm21, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm8
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm7, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm0, %zmm21
; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm10, %zmm6, %zmm17
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm16, %zmm6, %zmm24
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm13, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm15, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm26, %zmm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm23, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm26, %zmm10
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm6, %zmm12
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm10, %zmm6, %zmm5
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm8, %zmm0, %zmm17
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm31, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm29, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm7, %zmm4
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm27, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm25, %zmm8
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm7, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm0, %zmm26
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm8, %zmm0, %zmm29
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm4
+; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm0 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm7 = mem[0,1,1,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm7 = mem[2,1,3,3]
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm8 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm8 = mem[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm27 = mem[0,0,1,3]
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm24 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm11 = mem[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm12 = mem[0,2,2,3]
+; AVX512F-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm25 = mem[0,1,1,3]
+; AVX512F-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm15 = mem[2,1,3,3]
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm16 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm28 = mem[0,0,2,1]
+; AVX512F-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm18 = mem[0,0,1,3]
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm19 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm14 = mem[0,2,3,3,4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm7
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm12, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm12
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm12
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm7, %zmm0, %zmm15
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm12, %zmm0, %zmm22
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm8, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm24, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm23, %zmm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm28, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm19, %zmm7
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm23, %zmm7
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm8
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm7, %zmm0, %zmm10
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm5
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm13
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm24, 192(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 128(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, 256(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm22, 192(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm29, 128(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 64(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm28, (%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm27, 448(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm12, 704(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, 640(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm11, 576(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 512(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, 384(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, 768(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm19, 832(%rax)
-; AVX512F-FAST-NEXT: addq $2392, %rsp # imm = 0x958
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, (%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm21, 448(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 704(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm15, 640(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm26, 576(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 512(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm13, 384(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm20, 768(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, 832(%rax)
+; AVX512F-FAST-NEXT: addq $2200, %rsp # imm = 0x898
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
index 5d575ba3cec5..300a4c5aecc8 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
@@ -3482,506 +3482,206 @@ define void @store_i16_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
-; AVX512F-ONLY-FAST-LABEL: store_i16_stride8_vf32:
-; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $1704, %rsp # imm = 0x6A8
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r10), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r10), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <2,2,2,2,u,u,3,3>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = <0,0,0,0,u,u,1,1>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <0,u,0,u,u,u,1,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = <2,2,3,3,3,3,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = <0,0,1,1,1,1,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = <2,u,3,u,3,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %ymm18, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = <0,u,1,u,1,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm27, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %ymm18, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm17, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm16, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm20, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm15[0],xmm9[1],xmm15[1],xmm9[2],xmm15[2],xmm9[3],xmm15[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm16, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm17, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm17, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm26, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm24, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm26, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm27, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm26, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm24, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm27, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm24, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm26, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm26, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm27, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r10), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[8],ymm0[8],ymm4[9],ymm0[9],ymm4[10],ymm0[10],ymm4[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[12],ymm0[12],ymm4[13],ymm0[13],ymm4[14],ymm0[14],ymm4[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r10), %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm0[0],ymm14[0],ymm0[1],ymm14[1],ymm0[2],ymm14[2],ymm0[3],ymm14[3],ymm0[8],ymm14[8],ymm0[9],ymm14[9],ymm0[10],ymm14[10],ymm0[11],ymm14[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm14[4],ymm0[5],ymm14[5],ymm0[6],ymm14[6],ymm0[7],ymm14[7],ymm0[12],ymm14[12],ymm0[13],ymm14[13],ymm0[14],ymm14[14],ymm0[15],ymm14[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [0,2,4,6,4,6,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm14, %ymm30
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,0,4,4,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm1, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm14, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm14, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm14, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [2,1,6,5,6,5,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm3, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,4,5,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm1, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm3, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm1, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm3, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm1, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm31
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm29
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [4,6,2,3,6,7,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm3, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [4,4,2,1,6,5,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm1, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm3, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm1, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm3, %ymm19
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm1, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [6,5,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm3, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [4,5,1,3,5,7,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm1, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm3, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm1, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm3, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm1, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: movw $-30584, %ax # imm = 0x8888
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: movw $8738, %ax # imm = 0x2222
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm5 {%k2}
-; AVX512F-ONLY-FAST-NEXT: movb $-86, %al
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm5 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm4 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm4 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, (%rsp), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm1 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm1 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm0 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm28, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm6 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm3
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm16, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm7 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm7 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm26, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm8 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm8 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm29, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm9 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm9 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, 448(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $1704, %rsp # imm = 0x6A8
-; AVX512F-ONLY-FAST-NEXT: vzeroupper
-; AVX512F-ONLY-FAST-NEXT: retq
-;
-; AVX512DQ-FAST-LABEL: store_i16_stride8_vf32:
-; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $40, %rsp
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm13
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <2,2,2,2,u,u,3,3>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm7, %ymm26
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <0,u,0,u,u,u,1,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm20
-; AVX512DQ-FAST-NEXT: vmovdqa (%r10), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r10), %xmm11
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %xmm10
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm7, %ymm22
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <0,0,0,0,u,u,1,1>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm17, %ymm23
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <2,u,3,u,3,u,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm12, %ymm24
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = <0,u,1,u,1,u,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm21, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm9[4],xmm12[5],xmm9[5],xmm12[6],xmm9[6],xmm12[7],xmm9[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = <2,2,3,3,3,3,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm27
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = <0,0,1,1,1,1,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm28
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm29
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm30
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm31
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm3
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm20, %zmm20
-; AVX512DQ-FAST-NEXT: movw $-30584, %r11w # imm = 0x8888
-; AVX512DQ-FAST-NEXT: kmovw %r11d, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm22, %zmm23, %zmm20 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm25, %zmm23
-; AVX512DQ-FAST-NEXT: movw $8738, %r11w # imm = 0x2222
-; AVX512DQ-FAST-NEXT: kmovw %r11d, %k2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm27, %zmm28, %zmm23 {%k2}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm13[4],xmm14[4],xmm13[5],xmm14[5],xmm13[6],xmm14[6],xmm13[7],xmm14[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm7, %ymm14
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm16, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm7, %ymm11
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm17, %ymm10
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm18, %ymm15
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm21, %ymm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm30, %zmm24
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm31, %zmm6, %zmm24 {%k1}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm3, %zmm0, %zmm2 {%k2}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm29
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm11, %zmm10, %zmm29 {%k1}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm8, %zmm3
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm18, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm19, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm3 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm16
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm16 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm7
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm9[0],xmm12[1],xmm9[1],xmm12[2],xmm9[2],xmm12[3],xmm9[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm7 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[1],ymm1[1],ymm15[2],ymm1[2],ymm15[3],ymm1[3],ymm15[8],ymm1[8],ymm15[9],ymm1[9],ymm15[10],ymm1[10],ymm15[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [2,1,6,5,6,5,7,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm22, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r10), %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [0,2,4,6,4,6,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm25, %ymm17
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [0,0,4,4,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm26, %ymm18
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [6,5,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm14, %ymm28, %ymm19
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = [4,5,1,3,5,7,5,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm14, %ymm30, %ymm27
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm5[0],ymm12[0],ymm5[1],ymm12[1],ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[8],ymm12[8],ymm5[9],ymm12[9],ymm5[10],ymm12[10],ymm5[11],ymm12[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [4,6,2,3,6,7,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm20, %ymm31
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [4,4,2,1,6,5,6,5]
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm14, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm15[4],ymm1[4],ymm15[5],ymm1[5],ymm15[6],ymm1[6],ymm15[7],ymm1[7],ymm15[12],ymm1[12],ymm15[13],ymm1[13],ymm15[14],ymm1[14],ymm15[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm22, %ymm1
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm17, %zmm18, %zmm4 {%k1}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm27, %zmm15
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm31, %zmm13, %zmm15 {%k2}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm17
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm26, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm17 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm28, %ymm18
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm30, %ymm19
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm5[4],ymm12[4],ymm5[5],ymm12[5],ymm5[6],ymm12[6],ymm5[7],ymm12[7],ymm5[12],ymm12[12],ymm5[13],ymm12[13],ymm5[14],ymm12[14],ymm5[15],ymm12[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm27
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm14, %ymm31
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm21, %ymm9
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm22, %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa (%r10), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm11
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm25, %ymm13
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm26, %ymm12
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm21, %ymm21
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm22, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm10 = ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm1[0],ymm6[0],ymm1[1],ymm6[1],ymm1[2],ymm6[2],ymm1[3],ymm6[3],ymm1[8],ymm6[8],ymm1[9],ymm6[9],ymm1[10],ymm6[10],ymm1[11],ymm6[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm28, %ymm22
-; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm30, %ymm11
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm25, %ymm25
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm26, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm6[4],ymm1[5],ymm6[5],ymm1[6],ymm6[6],ymm1[7],ymm6[7],ymm1[12],ymm6[12],ymm1[13],ymm6[13],ymm1[14],ymm6[14],ymm1[15],ymm6[15]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm6
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm28, %ymm26
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm30, %ymm28
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[8],ymm0[8],ymm6[9],ymm0[9],ymm6[10],ymm0[10],ymm6[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm0[4],ymm6[5],ymm0[5],ymm6[6],ymm0[6],ymm6[7],ymm0[7],ymm6[12],ymm0[12],ymm6[13],ymm0[13],ymm6[14],ymm0[14],ymm6[15],ymm0[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm6
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm14, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm20
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm14, %ymm0
-; AVX512DQ-FAST-NEXT: movb $-86, %al
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k3
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, %zmm23 {%k3}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, %zmm2 {%k3}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm29, %zmm3 {%k3}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, %zmm7 {%k3}
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm15 {%k3}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm19, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm27, %zmm31, %zmm4 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, %zmm4 {%k3}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm13, %zmm12, %zmm8 {%k1}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm11, %zmm9
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm1, %zmm9 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm9 {%k3}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm5, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm25, %zmm10, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm28, %zmm5
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm20, %zmm0, %zmm5 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, %zmm5 {%k3}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 448(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, 384(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm23, 64(%rax)
-; AVX512DQ-FAST-NEXT: addq $40, %rsp
-; AVX512DQ-FAST-NEXT: vzeroupper
-; AVX512DQ-FAST-NEXT: retq
+; AVX512F-FAST-LABEL: store_i16_stride8_vf32:
+; AVX512F-FAST: # %bb.0:
+; AVX512F-FAST-NEXT: subq $552, %rsp # imm = 0x228
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512F-FAST-NEXT: vmovdqa (%r10), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %xmm2
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm26
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm1, %xmm27
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm20
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm1, %xmm21
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm18
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm1, %xmm19
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%r10), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm22
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm4
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm24
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm23
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm7
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm7[0],ymm8[1],ymm7[1],ymm8[2],ymm7[2],ymm8[3],ymm7[3],ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm25
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%r10), %xmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm16
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %xmm2
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm17
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm13
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm8[4],ymm7[4],ymm8[5],ymm7[5],ymm8[6],ymm7[6],ymm8[7],ymm7[7],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqa (%r10), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm11 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm26, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm26
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm27
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm28
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm15, %zmm15, %zmm31
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm10, %zmm10, %zmm29
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm8, %zmm8, %zmm30
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm10 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm20, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm21, %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm18, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm19, %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-FAST-NEXT: movw $-30584, %ax # imm = 0x8888
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
+; AVX512F-FAST-NEXT: vpermd %zmm24, %zmm0, %zmm24
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
+; AVX512F-FAST-NEXT: vpermd %zmm22, %zmm1, %zmm24 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm17, %zmm0, %zmm17
+; AVX512F-FAST-NEXT: vpermd %zmm16, %zmm1, %zmm17 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm16
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm22
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm13
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm14, %zmm14, %zmm14
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm12, %zmm12
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm7, %zmm7
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm11, %zmm11
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm9, %zmm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm5, %zmm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm2, %zmm2
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm10, %zmm10, %zmm10
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm8, %zmm8, %zmm8
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm15, %zmm15, %zmm15
+; AVX512F-FAST-NEXT: vpermd %zmm7, %zmm0, %zmm7
+; AVX512F-FAST-NEXT: vpermd %zmm12, %zmm1, %zmm7 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: movw $8738, %ax # imm = 0x2222
+; AVX512F-FAST-NEXT: kmovw %eax, %k2
+; AVX512F-FAST-NEXT: vpermd %zmm11, %zmm1, %zmm0 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
+; AVX512F-FAST-NEXT: vpermd %zmm25, %zmm1, %zmm9
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
+; AVX512F-FAST-NEXT: vpermd %zmm23, %zmm11, %zmm9 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm27, %zmm1, %zmm12
+; AVX512F-FAST-NEXT: vpermd %zmm26, %zmm11, %zmm12 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm3, %zmm1, %zmm3
+; AVX512F-FAST-NEXT: vpermd %zmm6, %zmm11, %zmm3 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm4, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vpermd %zmm5, %zmm11, %zmm1 {%k2}
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = <0,u,0,u,u,u,1,u,10,10,10,10,u,u,11,11>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,0,0,0,u,u,1,1,10,10,10,10,u,u,11,11>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm5 {%k1} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd %zmm31, %zmm4, %zmm11
+; AVX512F-FAST-NEXT: vpermd %zmm28, %zmm6, %zmm11 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm22, %zmm4, %zmm18
+; AVX512F-FAST-NEXT: vpermd %zmm16, %zmm6, %zmm18 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm10, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vpermd %zmm2, %zmm6, %zmm4 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = <0,u,1,u,1,u,u,u,10,u,11,u,11,u,u,u>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,0,1,1,1,1,u,u,10,10,11,11,11,11,u,u>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm2 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = <0,u,1,u,1,u,u,u,10,10,11,11,11,11,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm30, %zmm10, %zmm16
+; AVX512F-FAST-NEXT: vpermd %zmm29, %zmm6, %zmm16 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm14, %zmm10, %zmm14
+; AVX512F-FAST-NEXT: vpermd %zmm13, %zmm6, %zmm14 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm15, %zmm10, %zmm10
+; AVX512F-FAST-NEXT: vpermd %zmm8, %zmm6, %zmm10 {%k2}
+; AVX512F-FAST-NEXT: movb $-86, %al
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, %zmm2 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, %zmm12 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm11, %zmm16 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm18, %zmm14 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, %zmm3 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, %zmm1 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, %zmm10 {%k1}
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, (%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 192(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm16, 256(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm12, 448(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, 384(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 64(%rax)
+; AVX512F-FAST-NEXT: addq $552, %rsp # imm = 0x228
+; AVX512F-FAST-NEXT: vzeroupper
+; AVX512F-FAST-NEXT: retq
;
; AVX512BW-LABEL: store_i16_stride8_vf32:
; AVX512BW: # %bb.0:
@@ -7421,1073 +7121,381 @@ define void @store_i16_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
-; AVX512F-ONLY-FAST-LABEL: store_i16_stride8_vf64:
-; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $4968, %rsp # imm = 0x1368
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r10), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r10), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rax), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm2, %xmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = <2,2,2,2,u,u,3,3>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <0,0,0,0,u,u,1,1>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm21
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <0,u,0,u,u,u,1,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r10), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rax), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm18, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm16, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm18, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm17, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm2[0],xmm15[1],xmm2[1],xmm15[2],xmm2[2],xmm15[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm15[4],xmm2[4],xmm15[5],xmm2[5],xmm15[6],xmm2[6],xmm15[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r10), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm19, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm21, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm18, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm17, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm13, %ymm18, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm13, %ymm17, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm2[4],xmm13[5],xmm2[5],xmm13[6],xmm2[6],xmm13[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = <2,2,3,3,3,3,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm13, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %ymm16, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <0,0,1,1,1,1,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm16, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <2,u,3,u,3,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <0,u,1,u,1,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm22, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm22, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm14, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm14, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm13, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r10), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rax), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r10), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rax), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm13 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [0,2,4,6,4,6,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [0,0,4,4,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [2,1,6,5,6,5,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm19, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,4,5,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r10), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm11[0],ymm2[0],ymm11[1],ymm2[1],ymm11[2],ymm2[2],ymm11[3],ymm2[3],ymm11[8],ymm2[8],ymm11[9],ymm2[9],ymm11[10],ymm2[10],ymm11[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm11 = ymm11[4],ymm2[4],ymm11[5],ymm2[5],ymm11[6],ymm2[6],ymm11[7],ymm2[7],ymm11[12],ymm2[12],ymm11[13],ymm2[13],ymm11[14],ymm2[14],ymm11[15],ymm2[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm0, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm19, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm7[0],ymm2[0],ymm7[1],ymm2[1],ymm7[2],ymm2[2],ymm7[3],ymm2[3],ymm7[8],ymm2[8],ymm7[9],ymm2[9],ymm7[10],ymm2[10],ymm7[11],ymm2[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm7[4],ymm2[4],ymm7[5],ymm2[5],ymm7[6],ymm2[6],ymm7[7],ymm2[7],ymm7[12],ymm2[12],ymm7[13],ymm2[13],ymm7[14],ymm2[14],ymm7[15],ymm2[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm0, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm1, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm0, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm1, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm7[0],ymm5[0],ymm7[1],ymm5[1],ymm7[2],ymm5[2],ymm7[3],ymm5[3],ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm7[4],ymm5[4],ymm7[5],ymm5[5],ymm7[6],ymm5[6],ymm7[7],ymm5[7],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm0, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm1, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm0, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm1, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r10), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm11[0],ymm9[0],ymm11[1],ymm9[1],ymm11[2],ymm9[2],ymm11[3],ymm9[3],ymm11[8],ymm9[8],ymm11[9],ymm9[9],ymm11[10],ymm9[10],ymm11[11],ymm9[11]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm14, %ymm0, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm14, %ymm1, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm11[4],ymm9[4],ymm11[5],ymm9[5],ymm11[6],ymm9[6],ymm11[7],ymm9[7],ymm11[12],ymm9[12],ymm11[13],ymm9[13],ymm11[14],ymm9[14],ymm11[15],ymm9[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm1, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm19, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm13, %ymm19, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm13, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm19, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm15, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm19, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm19, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [4,6,2,3,6,7,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm17, %ymm4, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [4,4,2,1,6,5,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm17, %ymm0, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm21, %ymm4, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm21, %ymm0, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm26, %ymm4, %ymm29
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm26, %ymm0, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm27, %ymm4, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm27, %ymm0, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm0, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm0, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm0, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[8],ymm1[8],ymm6[9],ymm1[9],ymm6[10],ymm1[10],ymm6[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[12],ymm1[12],ymm6[13],ymm1[13],ymm6[14],ymm1[14],ymm6[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [6,5,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm6, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [4,5,1,3,5,7,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm0, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm16, %ymm6, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm16, %ymm0, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm24, %ymm6, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm24, %ymm0, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm25, %ymm6, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm25, %ymm0, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm6, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm12, %ymm0, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm6, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm0, %ymm19
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm6, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm0, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm6, %ymm31
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm0, %ymm30
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: movw $-30584, %ax # imm = 0x8888
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm9 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: movw $8738, %ax # imm = 0x2222
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm9 {%k2}
-; AVX512F-ONLY-FAST-NEXT: movb $-86, %al
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm9 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm8 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm8 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm8 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm7 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm7 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm7 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm6 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm6 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm5 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm5 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm4, %zmm3 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm1 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm1 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm0 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm0 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm18, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm10 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm10 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm22, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm13, %zmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm11 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm11 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm14, %zmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm13 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm13 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm14 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm14 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm12, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm12 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm12 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm19, %zmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm15 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm15 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm16 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm16 {%k3}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, (%rsp), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm17 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm17 {%k3}
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, 448(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 704(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, 640(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm11, 960(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 896(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 576(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 512(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, 832(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 768(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $4968, %rsp # imm = 0x1368
-; AVX512F-ONLY-FAST-NEXT: vzeroupper
-; AVX512F-ONLY-FAST-NEXT: retq
-;
-; AVX512DQ-FAST-LABEL: store_i16_stride8_vf64:
-; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $1720, %rsp # imm = 0x6B8
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <2,2,2,2,u,u,3,3>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = <0,u,0,u,u,u,1,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm14, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r10), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <0,0,0,0,u,u,1,1>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <2,u,3,u,3,u,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm1, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <0,u,1,u,1,u,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm16, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = <2,2,3,3,3,3,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <0,0,1,1,1,1,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm10, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm14, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r10), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rax), %xmm4
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm11
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm10, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm15, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm20, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm16, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm20, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm17, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r10), %xmm12
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rax), %xmm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm10, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm14, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm16, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm20, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm17, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm10, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm14, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm10, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm15, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm20, %ymm31
-; AVX512DQ-FAST-NEXT: vpermd %ymm13, %ymm16, %ymm28
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm29
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm17, %ymm26
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm10, %ymm27
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm14, %ymm24
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm10, %ymm25
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm15, %ymm23
-; AVX512DQ-FAST-NEXT: vmovdqa %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm20, %ymm22
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm16, %ymm21
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm19
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm17, %ymm18
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm17, %ymm30
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm7
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm10, %ymm17
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm14, %ymm16
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r10), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %xmm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm10, %ymm15
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm0, %ymm14
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm12
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm11
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm5, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm6
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm30, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: movw $-30584, %r11w # imm = 0x8888
-; AVX512DQ-FAST-NEXT: kmovw %r11d, %k2
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k2} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: movw $8738, %r11w # imm = 0x2222
-; AVX512DQ-FAST-NEXT: kmovw %r11d, %k1
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k1} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k2} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k1} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k2} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k1} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm0 {%k2} # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm28, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm29, %zmm26, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm24, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm25, %zmm23, %zmm0 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm19, %zmm18, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm16, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm14, %zmm0 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm5, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <0,u,0,u,u,u,1,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm8, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm10, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <0,0,0,0,u,u,1,1>
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm6, %ymm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm20, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <0,u,1,u,1,u,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm7, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm5 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm18
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <0,0,1,1,1,1,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm18 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm8, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm10, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm2 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm7, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm15
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm3, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm15 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [2,1,6,5,6,5,7,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm26, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm19, %ymm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r10), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rax), %ymm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [0,2,4,6,4,6,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm20, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [0,0,4,4,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm21, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm4, %zmm6 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [6,5,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm22, %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [4,5,1,3,5,7,5,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm23, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm17
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [4,6,2,3,6,7,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm24, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [4,4,2,1,6,5,6,5]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm25, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm4, %zmm17 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm4
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm4 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm22, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm23, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm27
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm27 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm26, %ymm3
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm19, %ymm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm28
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r10), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rax), %ymm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm20, %ymm6
-; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm21, %ymm5
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm5, %zmm28 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm22, %ymm9
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm23, %ymm7
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm29
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm24, %ymm16
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm25, %ymm10
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm16, %zmm10, %zmm29 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm30
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm30 {%k2}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm22, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm23, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm31
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm31 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm26, %ymm3
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm19, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm16
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r10), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm20, %ymm7
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm21, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm1, %zmm16 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm22, %ymm10
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm23, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm10[0],ymm3[1],ymm10[1],ymm3[2],ymm10[2],ymm3[3],ymm10[3],ymm3[8],ymm10[8],ymm3[9],ymm10[9],ymm3[10],ymm10[10],ymm3[11],ymm10[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm13 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm2[4],ymm4[4],ymm2[5],ymm4[5],ymm2[6],ymm4[6],ymm2[7],ymm4[7],ymm2[12],ymm4[12],ymm2[13],ymm4[13],ymm2[14],ymm4[14],ymm2[15],ymm4[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm14
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm20, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm21, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm14 {%k2}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm22, %ymm1
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm23, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm10[4],ymm3[5],ymm10[5],ymm3[6],ymm10[6],ymm3[7],ymm10[7],ymm3[12],ymm10[12],ymm3[13],ymm10[13],ymm3[14],ymm10[14],ymm3[15],ymm10[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm24, %ymm3
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm25, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm3, %zmm1, %zmm6 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm8
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm8[0],ymm11[0],ymm8[1],ymm11[1],ymm8[2],ymm11[2],ymm8[3],ymm11[3],ymm8[8],ymm11[8],ymm8[9],ymm11[9],ymm8[10],ymm11[10],ymm8[11],ymm11[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm26, %ymm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm19, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqa (%r10), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm20, %ymm9
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm21, %ymm7
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm9, %zmm7, %zmm12 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm22, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm23, %ymm10
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm10[0],ymm3[1],ymm10[1],ymm3[2],ymm10[2],ymm3[3],ymm10[3],ymm3[8],ymm10[8],ymm3[9],ymm10[9],ymm3[10],ymm10[10],ymm3[11],ymm10[11]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm2
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm0, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm8[4],ymm11[4],ymm8[5],ymm11[5],ymm8[6],ymm11[6],ymm8[7],ymm11[7],ymm8[12],ymm11[12],ymm8[13],ymm11[13],ymm8[14],ymm11[14],ymm8[15],ymm11[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm26, %ymm2
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm19, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm20, %ymm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm21, %ymm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm4, %zmm0 {%k2}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm22, %ymm4
-; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm23, %ymm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm3[4],ymm10[4],ymm3[5],ymm10[5],ymm3[6],ymm10[6],ymm3[7],ymm10[7],ymm3[12],ymm10[12],ymm3[13],ymm10[13],ymm3[14],ymm10[14],ymm3[15],ymm10[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm24, %ymm5
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm25, %ymm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm3, %zmm2 {%k1}
-; AVX512DQ-FAST-NEXT: movb $-86, %al
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm3 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, %zmm5 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, %zmm8 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm9 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm18 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm15 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm17 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm27 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm28, %zmm29 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm30, %zmm31 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, %zmm6 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 448(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, 384(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, 704(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm29, 640(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, 960(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, 896(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, 576(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 512(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 832(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 768(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 64(%rax)
-; AVX512DQ-FAST-NEXT: addq $1720, %rsp # imm = 0x6B8
-; AVX512DQ-FAST-NEXT: vzeroupper
-; AVX512DQ-FAST-NEXT: retq
+; AVX512F-FAST-LABEL: store_i16_stride8_vf64:
+; AVX512F-FAST: # %bb.0:
+; AVX512F-FAST-NEXT: subq $2504, %rsp # imm = 0x9C8
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512F-FAST-NEXT: vmovdqa (%r10), %xmm11
+; AVX512F-FAST-NEXT: vmovdqa 64(%r10), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%r10), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 96(%rax), %ymm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm2
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %ymm3
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %ymm4
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %ymm5
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %ymm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[8],ymm6[8],ymm8[9],ymm6[9],ymm8[10],ymm6[10],ymm8[11],ymm6[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm9, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[12],ymm6[12],ymm8[13],ymm6[13],ymm8[14],ymm6[14],ymm8[15],ymm6[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%r10), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 96(%rax), %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm2, %zmm2
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%r9), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %xmm3
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm5
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %xmm6
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 64(%r10), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa 64(%rax), %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %ymm4
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %ymm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %ymm8
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %ymm4
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm4[0],ymm8[0],ymm4[1],ymm8[1],ymm4[2],ymm8[2],ymm4[3],ymm8[3],ymm4[8],ymm8[8],ymm4[9],ymm8[9],ymm4[10],ymm8[10],ymm4[11],ymm8[11]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm8[4],ymm4[5],ymm8[5],ymm4[6],ymm8[6],ymm4[7],ymm8[7],ymm4[12],ymm8[12],ymm4[13],ymm8[13],ymm4[14],ymm8[14],ymm4[15],ymm8[15]
+; AVX512F-FAST-NEXT: vmovdqa 64(%rax), %xmm8
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm6, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 64(%r9), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %xmm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm31
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %xmm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm5, %zmm5, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm8, %zmm8, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm7, %zmm7, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%r10), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %ymm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm30
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm28
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm6, %zmm29
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm7
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm8, %zmm27
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm26
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm25
+; AVX512F-FAST-NEXT: vmovdqa 32(%r10), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rax), %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm21
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm19
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm5, %zmm5, %zmm24
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm3, %zmm3, %zmm23
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm22
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm20
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%r10), %ymm7
+; AVX512F-FAST-NEXT: vmovdqa (%rax), %ymm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[8],ymm7[8],ymm6[9],ymm7[9],ymm6[10],ymm7[10],ymm6[11],ymm7[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm18
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm4
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm15, %zmm16
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm14, %zmm17
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm13, %zmm15
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[12],ymm7[12],ymm6[13],ymm7[13],ymm6[14],ymm7[14],ymm6[15],ymm7[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm6, %zmm12
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm10
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm11[0],xmm1[1],xmm11[1],xmm1[2],xmm11[2],xmm1[3],xmm11[3]
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm7
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm11
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm3, %zmm3, %zmm9
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: movw $-30584, %ax # imm = 0x8888
+; AVX512F-FAST-NEXT: kmovw %eax, %k2
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,1,4,5,4,5,5,7,10,9,14,13,14,13,15,15]
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,4,4,4,4,6,5,8,10,12,14,12,14,14,15]
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm3 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vpermd %zmm31, %zmm1, %zmm31
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm31 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd %zmm28, %zmm1, %zmm28
+; AVX512F-FAST-NEXT: vpermd %zmm30, %zmm0, %zmm28 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm25, %zmm1, %zmm25
+; AVX512F-FAST-NEXT: vpermd %zmm26, %zmm0, %zmm25 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm16, %zmm1, %zmm16
+; AVX512F-FAST-NEXT: vpermd %zmm18, %zmm0, %zmm16 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm10, %zmm1, %zmm3
+; AVX512F-FAST-NEXT: vpermd %zmm12, %zmm0, %zmm3 {%k2}
+; AVX512F-FAST-NEXT: movw $8738, %ax # imm = 0x2222
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [4,5,1,3,5,7,5,7,14,13,11,11,15,15,15,15]
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm10 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [4,4,2,1,6,5,6,5,12,14,10,11,14,15,14,15]
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm10 {%k1} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm18 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm18 {%k1} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm26 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm26 {%k1} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm30 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm30 {%k1} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd %zmm27, %zmm0, %zmm27
+; AVX512F-FAST-NEXT: vpermd %zmm29, %zmm12, %zmm27 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm19, %zmm0, %zmm19
+; AVX512F-FAST-NEXT: vpermd %zmm21, %zmm12, %zmm19 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm15, %zmm0, %zmm15
+; AVX512F-FAST-NEXT: vpermd %zmm17, %zmm12, %zmm15 {%k1}
+; AVX512F-FAST-NEXT: vpermd %zmm7, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermd %zmm5, %zmm12, %zmm0 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = <0,u,0,u,u,u,1,u,10,10,10,10,u,u,11,11>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm7 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = <0,0,0,0,u,u,1,1,10,10,10,10,u,u,11,11>
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm7 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm17 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm17 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm21 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm21 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm29 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm29 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm1 {%k2} # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermd %zmm23, %zmm5, %zmm23
+; AVX512F-FAST-NEXT: vpermd %zmm24, %zmm12, %zmm23 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm20, %zmm5, %zmm20
+; AVX512F-FAST-NEXT: vpermd %zmm22, %zmm12, %zmm20 {%k2}
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm5, %zmm5
+; AVX512F-FAST-NEXT: vpermd %zmm11, %zmm12, %zmm5 {%k2}
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm8, %zmm8, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = <0,u,1,u,1,u,u,u,10,u,11,u,11,u,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm8, %zmm9, %zmm8
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm9
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = <0,0,1,1,1,1,u,u,10,10,11,11,11,11,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm8 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm12
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = <0,u,1,u,1,u,u,u,10,10,11,11,11,11,u,u>
+; AVX512F-FAST-NEXT: vpermd %zmm12, %zmm13, %zmm12
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm12 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm22
+; AVX512F-FAST-NEXT: vpermd %zmm22, %zmm13, %zmm22
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm22 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm24
+; AVX512F-FAST-NEXT: vpermd %zmm24, %zmm13, %zmm24
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm24 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm14
+; AVX512F-FAST-NEXT: vpermd %zmm14, %zmm13, %zmm14
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm14 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm6
+; AVX512F-FAST-NEXT: vpermd %zmm6, %zmm13, %zmm6
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm6 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vpermd %zmm4, %zmm13, %zmm4
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm11, %zmm4 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm9, %zmm9, %zmm9
+; AVX512F-FAST-NEXT: vpermd %zmm9, %zmm13, %zmm9
+; AVX512F-FAST-NEXT: vmovdqu64 (%rsp), %zmm13 # 64-byte Reload
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm13
+; AVX512F-FAST-NEXT: vpermd %zmm13, %zmm11, %zmm9 {%k1}
+; AVX512F-FAST-NEXT: movb $-86, %al
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, %zmm10 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, %zmm12 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm21, %zmm22 {%k1}
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm31, %zmm30 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm29, %zmm24 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, %zmm14 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm28, %zmm27 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm25, %zmm19 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm23, %zmm6 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm20, %zmm4 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm16, %zmm15 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, %zmm9 {%k1}
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, (%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm15, 128(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, 320(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm19, 448(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm27, 384(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, 576(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm24, 512(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm30, 704(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm26, 640(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm22, 832(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm12, 768(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm18, 960(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, 896(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512F-FAST-NEXT: addq $2504, %rsp # imm = 0x9C8
+; AVX512F-FAST-NEXT: vzeroupper
+; AVX512F-FAST-NEXT: retq
;
; AVX512BW-LABEL: store_i16_stride8_vf64:
; AVX512BW: # %bb.0:
@@ -8805,9 +7813,11 @@ define void @store_i16_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512BW-SLOW: {{.*}}
+; AVX512DQ-FAST: {{.*}}
; AVX512DQ-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
+; AVX512F-ONLY-FAST: {{.*}}
; AVX512F-ONLY-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index bf57ee77a355..81df38786284 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
@@ -947,12 +947,12 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-FAST-LABEL: store_i8_stride5_vf16:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm1
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm2
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm3
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm1
-; AVX512F-FAST-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm5
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm0
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm5
; AVX512F-FAST-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm6
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm6[0,2,0,2]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm7[0,8,u],zero,zero,ymm7[1,9,u],zero,zero,ymm7[2,10,u],zero,zero,ymm7[19,27,u],zero,zero,ymm7[20,28,u],zero,zero,ymm7[21,29,u],zero,zero
@@ -967,20 +967,18 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[2,6,u],zero,zero,ymm6[3,7,u],zero,zero,ymm6[8,12,u],zero,zero,ymm6[9,17,u],zero,zero,ymm6[22,18,u],zero,zero,ymm6[23,19,u],zero,zero,ymm6[24,28]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
; AVX512F-FAST-NEXT: vporq %zmm7, %zmm5, %zmm5
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <1,1,u,2,2,2,2,u>
-; AVX512F-FAST-NEXT: vpermd %ymm1, %ymm6, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <u,0,0,0,0,u,1,1>
-; AVX512F-FAST-NEXT: vpermd %ymm1, %ymm7, %ymm7
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,0,0,0,0,u,1,1,9,9,u,10,10,10,10,u>
+; AVX512F-FAST-NEXT: vpermd %zmm6, %zmm7, %zmm6
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u],zero,zero,xmm3[10,11,u],zero,zero,xmm3[12,13,u],zero,zero,xmm3[14,15,u]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,10,11],zero,zero,xmm0[u,12,13],zero,zero,xmm0[u,14,15],zero,zero,xmm0[u]
-; AVX512F-FAST-NEXT: vpor %xmm3, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[1,2,3,4],zero,xmm0[6,7,8,9],zero,xmm0[11,12,13,14],zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[12],zero,zero,zero,zero,xmm1[13],zero,zero,zero,zero,xmm1[14],zero,zero,zero,zero,xmm1[15]
-; AVX512F-FAST-NEXT: vpor %xmm1, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,10,11],zero,zero,xmm1[u,12,13],zero,zero,xmm1[u,14,15],zero,zero,xmm1[u]
+; AVX512F-FAST-NEXT: vpor %xmm3, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[1,2,3,4],zero,xmm1[6,7,8,9],zero,xmm1[11,12,13,14],zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[12],zero,zero,zero,zero,xmm0[13],zero,zero,zero,zero,xmm0[14],zero,zero,zero,zero,xmm0[15]
+; AVX512F-FAST-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512F-FAST-NEXT: vmovdqa %xmm0, 64(%r9)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, (%r9)
; AVX512F-FAST-NEXT: vzeroupper
@@ -1951,86 +1949,83 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-FAST-LABEL: store_i8_stride5_vf32:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm5
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm7
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9,u,11,u],zero,xmm7[10],zero,xmm7[12,u],zero
-; AVX512F-FAST-NEXT: vpor %xmm6, %xmm8, %xmm6
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm9 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm11 = xmm10[6],zero,xmm10[8],zero,xmm10[u,7],zero,xmm10[9],zero,xmm10[11],zero,xmm10[u,10],zero,xmm10[12],zero
-; AVX512F-FAST-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
-; AVX512F-FAST-NEXT: vpternlogq $226, %ymm6, %ymm11, %ymm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm6
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
+; AVX512F-FAST-NEXT: vpor %xmm5, %xmm7, %xmm5
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512F-FAST-NEXT: vpternlogq $226, %ymm8, %ymm7, %ymm5
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <1,1,u,2,2,2,2,u>
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm6, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,0,0,0,0,u,1,1>
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm8, %ymm8
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u]
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm7[6],zero,xmm7[8,u],zero,xmm7[7],zero,xmm7[9],zero,xmm7[11,u],zero,xmm7[10],zero,xmm7[12]
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm10 = xmm9[6],zero,xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[11],zero,xmm9[u,10],zero,xmm9[12],zero
+; AVX512F-FAST-NEXT: vpor %xmm8, %xmm10, %xmm8
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255]
+; AVX512F-FAST-NEXT: vpternlogq $226, %ymm5, %ymm10, %ymm8
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm5
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3],xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512F-FAST-NEXT: vpternlogq $226, %ymm7, %ymm6, %ymm4
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm4[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = <u,0,0,0,0,u,1,1,9,9,u,10,10,10,10,u>
+; AVX512F-FAST-NEXT: vpermd %zmm4, %zmm7, %zmm7
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u]
; AVX512F-FAST-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u,u],zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u,u,19]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u,u],zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u,u,19]
; AVX512F-FAST-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512F-FAST-NEXT: vpternlogq $226, %ymm5, %ymm11, %ymm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512F-FAST-NEXT: vpternlogq $226, %ymm5, %ymm10, %ymm8
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[21],zero,zero,ymm0[20],zero,ymm0[22],zero,ymm0[24],zero,zero,ymm0[23],zero,ymm0[25],zero,zero
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
; AVX512F-FAST-NEXT: vpternlogq $248, %ymm10, %ymm5, %ymm9
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm3[21],zero,ymm3[21,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm2[21],zero,ymm2[21,20],zero,ymm2[22],zero,ymm2[24],zero,ymm2[22,23],zero,ymm2[25]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
; AVX512F-FAST-NEXT: vpor %ymm5, %ymm11, %ymm5
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm5
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <4,u,5,5,5,5,u,6>
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm8, %ymm8
+; AVX512F-FAST-NEXT: vpermd %ymm4, %ymm8, %ymm8
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
; AVX512F-FAST-NEXT: vpandn %ymm8, %ymm9, %ymm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[12],zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,ymm0[18],zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[12],zero,zero,zero,zero,ymm4[13],zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,ymm4[18],zero
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm8
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,ymm3[26],zero,ymm3[28],zero,ymm3[30],zero,zero,ymm3[29],zero,ymm3[31],zero,zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512F-FAST-NEXT: vpternlogq $248, %ymm10, %ymm4, %ymm3
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero,zero
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512F-FAST-NEXT: vpternlogq $248, %ymm10, %ymm3, %ymm2
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm3, %ymm7, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <6,6,6,u,7,7,7,7>
-; AVX512F-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa %ymm0, 128(%r9)
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[26],zero,ymm0[28],zero,zero,ymm0[27],zero,ymm0[29],zero,ymm0[31],zero,zero,ymm0[30],zero
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm2, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <6,6,6,u,7,7,7,7>
+; AVX512F-FAST-NEXT: vpermd %ymm4, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa %ymm1, 128(%r9)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 64(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, (%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, (%r9)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
@@ -4086,180 +4081,177 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-FAST-LABEL: store_i8_stride5_vf64:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19>
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm3
-; AVX512F-FAST-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = <8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11>
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm1, %xmm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = <128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128>
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm6
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm4, %xmm25
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm3, %xmm18
-; AVX512F-FAST-NEXT: vpor %xmm1, %xmm6, %xmm1
-; AVX512F-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm7, %ymm8
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = <u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm13, %ymm14, %ymm10
-; AVX512F-FAST-NEXT: vpor %ymm8, %ymm10, %ymm3
-; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm5, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19>
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm1
+; AVX512F-FAST-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = <8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11>
+; AVX512F-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = <128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128>
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm2, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm19
+; AVX512F-FAST-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm8
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm9
+; AVX512F-FAST-NEXT: vpor %ymm8, %ymm9, %ymm7
+; AVX512F-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm8
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12>
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm8, %xmm11
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = <6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128>
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm10, %xmm12
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm4, %xmm26
-; AVX512F-FAST-NEXT: vporq %xmm11, %xmm12, %xmm19
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm22
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[26],zero,ymm0[28],zero,zero,ymm0[27],zero,ymm0[29],zero,ymm0[31],zero,zero,ymm0[30],zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[21],zero,zero,ymm0[20],zero,ymm0[22],zero,ymm0[24],zero,zero,ymm0[23],zero,ymm0[25],zero,zero
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm23
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,ymm14[26],zero,ymm14[28],zero,ymm14[30],zero,zero,ymm14[29],zero,ymm14[31],zero,zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[19],zero,ymm7[21],zero,zero,ymm7[20],zero,ymm7[22],zero,ymm7[24],zero,zero,ymm7[23],zero
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm20
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm6, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512F-FAST-NEXT: vpshufb %ymm13, %ymm4, %ymm13
-; AVX512F-FAST-NEXT: vporq %ymm2, %ymm13, %ymm21
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = <128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12>
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm8, %xmm12
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm7, %xmm26
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = <6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128>
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm9, %xmm13
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm7, %xmm27
+; AVX512F-FAST-NEXT: vporq %xmm12, %xmm13, %xmm20
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm5[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm22
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,ymm3[27],zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30],zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero,zero
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm23
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[27],zero,zero,ymm4[26],zero,ymm4[28],zero,ymm4[30],zero,zero,ymm4[29],zero,ymm4[31],zero,zero
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[19],zero,ymm0[21],zero,zero,ymm0[20],zero,ymm0[22],zero,ymm0[24],zero,zero,ymm0[23],zero
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm18
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm12
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm12, %ymm13
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm7
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm7, %ymm1
+; AVX512F-FAST-NEXT: vporq %ymm13, %ymm1, %ymm21
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm13
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm13, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm9
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm9, %ymm5
-; AVX512F-FAST-NEXT: vporq %ymm2, %ymm5, %ymm24
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm5
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm16
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm1, %xmm15
-; AVX512F-FAST-NEXT: vporq %xmm5, %xmm15, %xmm25
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm5
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm26, %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm3, %xmm11
-; AVX512F-FAST-NEXT: vporq %xmm5, %xmm11, %xmm29
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [1,1,2,2,2,2,2,2]
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm11
-; AVX512F-FAST-NEXT: vpermd %ymm11, %ymm26, %ymm27
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
-; AVX512F-FAST-NEXT: vpandnq %ymm27, %ymm28, %ymm27
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm12
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm27, %zmm31
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [4,6,5,5,5,5,4,6]
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm13, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm11
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm11, %ymm10
+; AVX512F-FAST-NEXT: vporq %ymm1, %ymm10, %ymm24
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm6, %xmm2, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm16
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm4
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm4, %xmm15
+; AVX512F-FAST-NEXT: vporq %xmm1, %xmm15, %xmm25
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm15
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm27, %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm5
+; AVX512F-FAST-NEXT: vporq %xmm15, %xmm5, %xmm26
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [1,1,2,2,2,2,2,2]
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm5
+; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm15, %ymm15
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
+; AVX512F-FAST-NEXT: vpandnq %ymm15, %ymm27, %ymm15
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm14
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm15, %zmm30
+; AVX512F-FAST-NEXT: vmovdqa64 (%r8), %zmm31
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = <4,u,5,5,5,5,u,6,30,30,30,u,31,31,31,31>
+; AVX512F-FAST-NEXT: vpermi2d %zmm31, %zmm5, %zmm27
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = <4,u,5,5,5,5,u,6>
; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm5
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm27, %ymm28
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX512F-FAST-NEXT: vpandnq %ymm28, %ymm30, %ymm28
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm2, %zmm28
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <9,u,11,u,u,10,u,12,u,14,u,u,13,u,15,u,9,u,11,u,u,10,u,12,u,14,u,u,13,u,15,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm7, %ymm12
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[19],zero,ymm6[21],zero,zero,ymm6[20],zero,ymm6[22],zero,ymm6[24],zero,zero,ymm6[23],zero
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm6, %ymm15
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25>
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm14, %ymm14
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm6
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[27],zero,zero,ymm4[26],zero,ymm4[28],zero,ymm4[30],zero,zero,ymm4[29],zero,ymm4[31],zero,zero
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm12[2,2,3,3]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm13[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm28, %ymm28
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX512F-FAST-NEXT: vpandnq %ymm28, %ymm29, %ymm28
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm3, %zmm28
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <9,u,11,u,u,10,u,12,u,14,u,u,13,u,15,u,9,u,11,u,u,10,u,12,u,14,u,u,13,u,15,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm14
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[19],zero,ymm12[21],zero,zero,ymm12[20],zero,ymm12[22],zero,ymm12[24],zero,zero,ymm12[23],zero
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm12, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25>
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm3
+; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm7, %ymm12
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,zero,ymm7[26],zero,ymm7[28],zero,ymm7[30],zero,zero,ymm7[29],zero,ymm7[31],zero,zero
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm14[2,2,3,3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm13[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm13 = ymm13[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm9[21],zero,zero,ymm9[20],zero,ymm9[22],zero,ymm9[24],zero,zero,ymm9[23],zero,ymm9[25],zero,zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm9[26],zero,ymm9[28],zero,zero,ymm9[27],zero,ymm9[29],zero,ymm9[31],zero,zero,ymm9[30],zero
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm18, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3],xmm7[4],xmm0[4],xmm7[5],xmm0[5],xmm7[6],xmm0[6],xmm7[7],xmm0[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm7
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3],xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
-; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm14[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm9 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[21],zero,zero,ymm11[20],zero,ymm11[22],zero,ymm11[24],zero,zero,ymm11[23],zero,ymm11[25],zero,zero
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm11[26],zero,ymm11[28],zero,zero,ymm11[27],zero,ymm11[29],zero,ymm11[31],zero,zero,ymm11[30],zero
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm19, %xmm10
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm15[0],xmm10[0],xmm15[1],xmm10[1],xmm15[2],xmm10[2],xmm15[3],xmm10[3],xmm15[4],xmm10[4],xmm15[5],xmm10[5],xmm15[6],xmm10[6],xmm15[7],xmm10[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm15[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm25, %zmm1, %zmm18
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm8, %xmm8
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm10
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm10 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm15, %xmm15
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm4, %xmm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm10 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm8, %xmm8
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm17, %xmm1
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm3, %xmm3
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm29, %zmm3, %zmm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [6,6,6,6,7,7,7,7]
-; AVX512F-FAST-NEXT: vpermd %ymm11, %ymm15, %ymm25
-; AVX512F-FAST-NEXT: vpermd %ymm11, %ymm27, %ymm27
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm29 = [0,0,0,0,0,0,1,1]
-; AVX512F-FAST-NEXT: vpermd %ymm11, %ymm29, %ymm11
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm15, %ymm15
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm26, %ymm26
-; AVX512F-FAST-NEXT: vpermd %ymm5, %ymm29, %ymm5
-; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm17 = mem[0,0,1,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm16 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm19[0,0,1,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm16, %zmm19, %zmm17
-; AVX512F-FAST-NEXT: vpor %ymm2, %ymm6, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm21, %zmm2
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,1,1]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm25, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3],xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm26, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm31, %zmm5
+; AVX512F-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: # ymm6 = mem[0,0,1,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm20[0,0,1,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 32-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm16, %zmm10
+; AVX512F-FAST-NEXT: vpor %ymm0, %ymm12, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512F-FAST-NEXT: vpternlogq $248, %ymm6, %ymm10, %ymm12
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm24, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm19, %zmm10
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm22[2,2,3,3,6,6,7,7]
+; AVX512F-FAST-NEXT: vpternlogq $248, %ymm6, %ymm14, %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm24, %zmm9
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm16, %zmm9
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm22[2,2,3,3,6,6,7,7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm12 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm12
-; AVX512F-FAST-NEXT: vpternlogq $248, %ymm6, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vpandq %ymm6, %ymm30, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm7, %zmm2
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm6 = zmm20[2,2,3,3,6,6,7,7]
-; AVX512F-FAST-NEXT: vporq %zmm6, %zmm2, %zmm2
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm12, %zmm6, %zmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm6, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm31
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm27, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm28
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm18[0,0,1,1,4,4,5,5]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm3[0,0,1,1,4,4,5,5]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm5, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm12
+; AVX512F-FAST-NEXT: vpternlogq $248, %ymm6, %ymm2, %ymm7
+; AVX512F-FAST-NEXT: vpandq %ymm6, %ymm29, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm18[2,2,3,3,6,6,7,7]
+; AVX512F-FAST-NEXT: vporq %zmm2, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm12, %zmm2, %zmm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm11
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm30
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm27
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm28
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = <6,6,6,u,7,7,7,7,u,8,8,8,8,u,9,9>
+; AVX512F-FAST-NEXT: vpermd %zmm31, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm4[0,0,1,1,4,4,5,5]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = <u,0,0,0,0,u,1,1,9,9,u,10,10,10,10,u>
+; AVX512F-FAST-NEXT: vpermd %zmm5, %zmm2, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
; AVX512F-FAST-NEXT: vmovdqa64 %zmm28, 64(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 128(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, 256(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm31, 192(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, (%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 128(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm27, 256(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm30, 192(%r9)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
index 26408c2ed15a..d4566c65cbcc 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
@@ -4587,472 +4587,465 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-SLOW-LABEL: store_i8_stride6_vf64:
; AVX512BW-SLOW: # %bb.0:
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %ymm14
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %ymm15
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm15[0],ymm14[0],ymm15[1],ymm14[1],ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[4],ymm14[4],ymm15[5],ymm14[5],ymm15[6],ymm14[6],ymm15[7],ymm14[7],ymm15[16],ymm14[16],ymm15[17],ymm14[17],ymm15[18],ymm14[18],ymm15[19],ymm14[19],ymm15[20],ymm14[20],ymm15[21],ymm14[21],ymm15[22],ymm14[22],ymm15[23],ymm14[23]
+; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512BW-SLOW-NEXT: vmovdqa 32(%rsi), %xmm8
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm10
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm12 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm12, %zmm5
+; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %xmm1
; AVX512BW-SLOW-NEXT: vmovdqa 32(%rcx), %xmm7
-; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm8
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm11, %ymm2
-; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %ymm14
-; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %ymm15
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm15[0],ymm14[0],ymm15[1],ymm14[1],ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[4],ymm14[4],ymm15[5],ymm14[5],ymm15[6],ymm14[6],ymm15[7],ymm14[7],ymm15[16],ymm14[16],ymm15[17],ymm14[17],ymm15[18],ymm14[18],ymm15[19],ymm14[19],ymm15[20],ymm14[20],ymm15[21],ymm14[21],ymm15[22],ymm14[22],ymm15[23],ymm14[23]
-; AVX512BW-SLOW-NEXT: vprold $16, %ymm3, %ymm3
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm4
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rsi), %ymm17
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdi), %ymm18
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm18[0],ymm17[0],ymm18[1],ymm17[1],ymm18[2],ymm17[2],ymm18[3],ymm17[3],ymm18[4],ymm17[4],ymm18[5],ymm17[5],ymm18[6],ymm17[6],ymm18[7],ymm17[7],ymm18[16],ymm17[16],ymm18[17],ymm17[17],ymm18[18],ymm17[18],ymm18[19],ymm17[19],ymm18[20],ymm17[20],ymm18[21],ymm17[21],ymm18[22],ymm17[22],ymm18[23],ymm17[23]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [8,11,10,9,8,11,10,9,8,11,10,9,12,13,14,13]
-; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm20, %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rsi), %xmm10
-; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm12
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm6, %ymm21, %ymm6
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm9
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm4, %ymm20, %ymm4
+; AVX512BW-SLOW-NEXT: vmovdqa64 (%rcx), %ymm17
+; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdx), %ymm18
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm18[0],ymm17[0],ymm18[1],ymm17[1],ymm18[2],ymm17[2],ymm18[3],ymm17[3],ymm18[4],ymm17[4],ymm18[5],ymm17[5],ymm18[6],ymm17[6],ymm18[7],ymm17[7],ymm18[16],ymm17[16],ymm18[17],ymm17[17],ymm18[18],ymm17[18],ymm18[19],ymm17[19],ymm18[20],ymm17[20],ymm18[21],ymm17[21],ymm18[22],ymm17[22],ymm18[23],ymm17[23]
+; AVX512BW-SLOW-NEXT: vprold $16, %ymm6, %ymm6
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
; AVX512BW-SLOW-NEXT: movl $613566756, %r10d # imm = 0x24924924
; AVX512BW-SLOW-NEXT: kmovd %r10d, %k3
; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm4, %zmm5 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm22
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm21
; AVX512BW-SLOW-NEXT: vmovdqa64 (%r8), %ymm16
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = <2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm16, %ymm4
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm22 = <2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm22, %ymm16, %ymm4
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512BW-SLOW-NEXT: movw $18724, %r10w # imm = 0x4924
; AVX512BW-SLOW-NEXT: kmovd %r10d, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm4, %ymm22 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm4, %ymm21 {%k1}
; AVX512BW-SLOW-NEXT: vmovdqa64 (%r9), %ymm19
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = <u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm19, %ymm4
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = <u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm19, %ymm4
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512BW-SLOW-NEXT: movl $545392672, %r10d # imm = 0x20820820
; AVX512BW-SLOW-NEXT: kmovd %r10d, %k2
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm4, %ymm22 {%k2}
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm4, %ymm21 {%k2}
; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm4
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%r8), %xmm9
+; AVX512BW-SLOW-NEXT: vmovdqa 32(%r8), %xmm11
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
; AVX512BW-SLOW-NEXT: movw $-28087, %r10w # imm = 0x9249
; AVX512BW-SLOW-NEXT: kmovd %r10d, %k4
-; AVX512BW-SLOW-NEXT: vpermw %ymm6, %ymm25, %ymm5 {%k4}
+; AVX512BW-SLOW-NEXT: vpermw %ymm6, %ymm24, %ymm5 {%k4}
; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %xmm6
; AVX512BW-SLOW-NEXT: vmovdqa 32(%r9), %xmm13
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm26 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm26, %ymm25, %ymm26
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm25, %ymm24, %ymm25
; AVX512BW-SLOW-NEXT: movl $-2113396606, %r10d # imm = 0x82082082
; AVX512BW-SLOW-NEXT: kmovd %r10d, %k7
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm26, %ymm5 {%k7}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm5, %zmm5
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm22, %ymm11, %ymm11
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm22
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm26
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm27 = ymm26[0],ymm22[0],ymm26[1],ymm22[1],ymm26[2],ymm22[2],ymm26[3],ymm22[3],ymm26[4],ymm22[4],ymm26[5],ymm22[5],ymm26[6],ymm22[6],ymm26[7],ymm22[7],ymm26[16],ymm22[16],ymm26[17],ymm22[17],ymm26[18],ymm22[18],ymm26[19],ymm22[19],ymm26[20],ymm22[20],ymm26[21],ymm22[21],ymm26[22],ymm22[22],ymm26[23],ymm22[23]
-; AVX512BW-SLOW-NEXT: vprold $16, %ymm27, %ymm27
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm27, %zmm11, %zmm11
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm27
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm28
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm27[0],ymm28[1],ymm27[1],ymm28[2],ymm27[2],ymm28[3],ymm27[3],ymm28[4],ymm27[4],ymm28[5],ymm27[5],ymm28[6],ymm27[6],ymm28[7],ymm27[7],ymm28[16],ymm27[16],ymm28[17],ymm27[17],ymm28[18],ymm27[18],ymm28[19],ymm27[19],ymm28[20],ymm27[20],ymm28[21],ymm27[21],ymm28[22],ymm27[22],ymm28[23],ymm27[23]
-; AVX512BW-SLOW-NEXT: vpermw %ymm29, %ymm20, %ymm20
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm29 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm29, %ymm21, %ymm21
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm21, %zmm20
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm11, %zmm20 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm20, %ymm11
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm21
-; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm21, %ymm23
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm25, %ymm5 {%k7}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm5, %zmm5
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm21
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm25
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm25[0],ymm21[0],ymm25[1],ymm21[1],ymm25[2],ymm21[2],ymm25[3],ymm21[3],ymm25[4],ymm21[4],ymm25[5],ymm21[5],ymm25[6],ymm21[6],ymm25[7],ymm21[7],ymm25[16],ymm21[16],ymm25[17],ymm21[17],ymm25[18],ymm21[18],ymm25[19],ymm21[19],ymm25[20],ymm21[20],ymm25[21],ymm21[21],ymm25[22],ymm21[22],ymm25[23],ymm21[23]
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm27 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm27, %zmm26
+; AVX512BW-SLOW-NEXT: vpermw %zmm26, %zmm12, %zmm12
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm26 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm26, %ymm20, %ymm20
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm26
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm27
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm26[0],ymm27[1],ymm26[1],ymm27[2],ymm26[2],ymm27[3],ymm26[3],ymm27[4],ymm26[4],ymm27[5],ymm26[5],ymm27[6],ymm26[6],ymm27[7],ymm26[7],ymm27[16],ymm26[16],ymm27[17],ymm26[17],ymm27[18],ymm26[18],ymm27[19],ymm26[19],ymm27[20],ymm26[20],ymm27[21],ymm26[21],ymm27[22],ymm26[22],ymm27[23],ymm26[23]
+; AVX512BW-SLOW-NEXT: vprold $16, %ymm28, %ymm28
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm20, %zmm20
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm20, %zmm12 {%k3}
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm12, %ymm20
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm28
+; AVX512BW-SLOW-NEXT: vpshufb %ymm22, %ymm28, %ymm22
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm22, %ymm20 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %ymm22
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm22, %ymm23
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm23, %ymm11 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %ymm23
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm23, %ymm24
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm24, %ymm11 {%k2}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm24, %ymm25, %ymm20 {%k4}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm24, %ymm25, %ymm24
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm24, %ymm20 {%k7}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm20, %zmm11
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = <8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm27, %ymm20
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm28, %ymm25
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm25[0],ymm20[0],ymm25[1],ymm20[1],ymm25[2],ymm20[2],ymm25[3],ymm20[3],ymm25[4],ymm20[4],ymm25[5],ymm20[5],ymm25[6],ymm20[6],ymm25[7],ymm20[7],ymm25[16],ymm20[16],ymm25[17],ymm20[17],ymm25[18],ymm20[18],ymm25[19],ymm20[19],ymm25[20],ymm20[20],ymm25[21],ymm20[21],ymm25[22],ymm20[22],ymm25[23],ymm20[23]
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm23, %ymm20 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm11[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm24, %ymm12 {%k4}
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm24, %ymm23
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm23, %ymm12 {%k7}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm12, %zmm12
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = <8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm21, %ymm20
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm25, %ymm24
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm24[0],ymm20[0],ymm24[1],ymm20[1],ymm24[2],ymm20[2],ymm24[3],ymm20[3],ymm24[4],ymm20[4],ymm24[5],ymm20[5],ymm24[6],ymm20[6],ymm24[7],ymm20[7],ymm24[16],ymm20[16],ymm24[17],ymm20[17],ymm24[18],ymm20[18],ymm24[19],ymm20[19],ymm24[20],ymm20[20],ymm24[21],ymm20[21],ymm24[22],ymm20[22],ymm24[23],ymm20[23]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm25 = ymm28[8],ymm27[8],ymm28[9],ymm27[9],ymm28[10],ymm27[10],ymm28[11],ymm27[11],ymm28[12],ymm27[12],ymm28[13],ymm27[13],ymm28[14],ymm27[14],ymm28[15],ymm27[15],ymm28[24],ymm27[24],ymm28[25],ymm27[25],ymm28[26],ymm27[26],ymm28[27],ymm27[27],ymm28[28],ymm27[28],ymm28[29],ymm27[29],ymm28[30],ymm27[30],ymm28[31],ymm27[31]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm25, %ymm27, %ymm25
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm20
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm25, %ymm22, %ymm28
-; AVX512BW-SLOW-NEXT: vpshufb %ymm25, %ymm26, %ymm29
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm22 = ymm26[8],ymm22[8],ymm26[9],ymm22[9],ymm26[10],ymm22[10],ymm26[11],ymm22[11],ymm26[12],ymm22[12],ymm26[13],ymm22[13],ymm26[14],ymm22[14],ymm26[15],ymm22[15],ymm26[24],ymm22[24],ymm26[25],ymm22[25],ymm26[26],ymm22[26],ymm26[27],ymm22[27],ymm26[28],ymm22[28],ymm26[29],ymm22[29],ymm26[30],ymm22[30],ymm26[31],ymm22[31]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm22, %ymm26, %ymm22
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm28, %zmm22
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm20, %zmm22 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm22, %ymm20
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm28 = <10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm28, %ymm21, %ymm29
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm21 = ymm25[8],ymm21[8],ymm25[9],ymm21[9],ymm25[10],ymm21[10],ymm25[11],ymm21[11],ymm25[12],ymm21[12],ymm25[13],ymm21[13],ymm25[14],ymm21[14],ymm25[15],ymm21[15],ymm25[24],ymm21[24],ymm25[25],ymm21[25],ymm25[26],ymm21[26],ymm25[27],ymm21[27],ymm25[28],ymm21[28],ymm25[29],ymm21[29],ymm25[30],ymm21[30],ymm25[31],ymm21[31]
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm21, %ymm24, %ymm21
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm20, %zmm20
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm26, %ymm25
+; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm27, %ymm29
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm29[0],ymm25[0],ymm29[1],ymm25[1],ymm29[2],ymm25[2],ymm29[3],ymm25[3],ymm29[4],ymm25[4],ymm29[5],ymm25[5],ymm29[6],ymm25[6],ymm29[7],ymm25[7],ymm29[16],ymm25[16],ymm29[17],ymm25[17],ymm29[18],ymm25[18],ymm29[19],ymm25[19],ymm29[20],ymm25[20],ymm29[21],ymm25[21],ymm29[22],ymm25[22],ymm29[23],ymm25[23]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm26 = ymm27[8],ymm26[8],ymm27[9],ymm26[9],ymm27[10],ymm26[10],ymm27[11],ymm26[11],ymm27[12],ymm26[12],ymm27[13],ymm26[13],ymm27[14],ymm26[14],ymm27[15],ymm26[15],ymm27[24],ymm26[24],ymm27[25],ymm26[25],ymm27[26],ymm26[26],ymm27[27],ymm26[27],ymm27[28],ymm26[28],ymm27[29],ymm26[29],ymm27[30],ymm26[30],ymm27[31],ymm26[31]
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm26, %ymm27, %ymm26
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm25
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm20, %zmm25 {%k3}
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm25, %ymm20
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = <10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm26, %ymm28, %ymm29
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm29[2,2,2,3]
; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm29, %ymm20 {%k4}
; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm29 = <u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm29, %ymm23, %ymm30
+; AVX512BW-SLOW-NEXT: vpshufb %ymm29, %ymm22, %ymm30
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm30, %ymm20 {%k7}
; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm30 = <6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm30, %ymm21, %ymm21
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vpshufb %ymm30, %ymm28, %ymm28
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
; AVX512BW-SLOW-NEXT: movw $9362, %cx # imm = 0x2492
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k5
-; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm21, %ymm22 {%k5}
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = <u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm23, %ymm23
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm28, %ymm25 {%k5}
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm28 = <u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u>
+; AVX512BW-SLOW-NEXT: vpshufb %ymm28, %ymm22, %ymm22
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
; AVX512BW-SLOW-NEXT: movl $136348168, %ecx # imm = 0x8208208
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k6
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm23, %ymm22 {%k6}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm22, %zmm20
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm17, %ymm22
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm18, %ymm23
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm22, %ymm25 {%k6}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm25, %zmm20
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm14, %ymm22
+; AVX512BW-SLOW-NEXT: vpshufb %ymm23, %ymm15, %ymm23
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm22 = ymm23[0],ymm22[0],ymm23[1],ymm22[1],ymm23[2],ymm22[2],ymm23[3],ymm22[3],ymm23[4],ymm22[4],ymm23[5],ymm22[5],ymm23[6],ymm22[6],ymm23[7],ymm22[7],ymm23[16],ymm22[16],ymm23[17],ymm22[17],ymm23[18],ymm22[18],ymm23[19],ymm22[19],ymm23[20],ymm22[20],ymm23[21],ymm22[21],ymm23[22],ymm22[22],ymm23[23],ymm22[23]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm14 = ymm15[8],ymm14[8],ymm15[9],ymm14[9],ymm15[10],ymm14[10],ymm15[11],ymm14[11],ymm15[12],ymm14[12],ymm15[13],ymm14[13],ymm15[14],ymm14[14],ymm15[15],ymm14[15],ymm15[24],ymm14[24],ymm15[25],ymm14[25],ymm15[26],ymm14[26],ymm15[27],ymm14[27],ymm15[28],ymm14[28],ymm15[29],ymm14[29],ymm15[30],ymm14[30],ymm15[31],ymm14[31]
+; AVX512BW-SLOW-NEXT: vpermw %ymm14, %ymm24, %ymm14
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm22, %zmm14
+; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm17, %ymm15
+; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm18, %ymm21
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm21[0],ymm15[0],ymm21[1],ymm15[1],ymm21[2],ymm15[2],ymm21[3],ymm15[3],ymm21[4],ymm15[4],ymm21[5],ymm15[5],ymm21[6],ymm15[6],ymm21[7],ymm15[7],ymm21[16],ymm15[16],ymm21[17],ymm15[17],ymm21[18],ymm15[18],ymm21[19],ymm15[19],ymm21[20],ymm15[20],ymm21[21],ymm15[21],ymm21[22],ymm15[22],ymm21[23],ymm15[23]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm17 = ymm18[8],ymm17[8],ymm18[9],ymm17[9],ymm18[10],ymm17[10],ymm18[11],ymm17[11],ymm18[12],ymm17[12],ymm18[13],ymm17[13],ymm18[14],ymm17[14],ymm18[15],ymm17[15],ymm18[24],ymm17[24],ymm18[25],ymm17[25],ymm18[26],ymm17[26],ymm18[27],ymm17[27],ymm18[28],ymm17[28],ymm18[29],ymm17[29],ymm18[30],ymm17[30],ymm18[31],ymm17[31]
; AVX512BW-SLOW-NEXT: vpermw %ymm17, %ymm27, %ymm17
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm22, %zmm17
-; AVX512BW-SLOW-NEXT: vpshufb %ymm25, %ymm14, %ymm18
-; AVX512BW-SLOW-NEXT: vpshufb %ymm25, %ymm15, %ymm22
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm22[0],ymm18[0],ymm22[1],ymm18[1],ymm22[2],ymm18[2],ymm22[3],ymm18[3],ymm22[4],ymm18[4],ymm22[5],ymm18[5],ymm22[6],ymm18[6],ymm22[7],ymm18[7],ymm22[16],ymm18[16],ymm22[17],ymm18[17],ymm22[18],ymm18[18],ymm22[19],ymm18[19],ymm22[20],ymm18[20],ymm22[21],ymm18[21],ymm22[22],ymm18[22],ymm22[23],ymm18[23]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm18[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm14 = ymm15[8],ymm14[8],ymm15[9],ymm14[9],ymm15[10],ymm14[10],ymm15[11],ymm14[11],ymm15[12],ymm14[12],ymm15[13],ymm14[13],ymm15[14],ymm14[14],ymm15[15],ymm14[15],ymm15[24],ymm14[24],ymm15[25],ymm14[25],ymm15[26],ymm14[26],ymm15[27],ymm14[27],ymm15[28],ymm14[28],ymm15[29],ymm14[29],ymm15[30],ymm14[30],ymm15[31],ymm14[31]
-; AVX512BW-SLOW-NEXT: vpermw %ymm14, %ymm26, %ymm14
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm18, %zmm14
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm17, %zmm14 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm14, %ymm15
-; AVX512BW-SLOW-NEXT: vpshufb %ymm28, %ymm16, %ymm17
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm15, %zmm15
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm14, %zmm15 {%k3}
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm15, %ymm14
+; AVX512BW-SLOW-NEXT: vpshufb %ymm26, %ymm16, %ymm17
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm17, %ymm15 {%k4}
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm17, %ymm14 {%k4}
; AVX512BW-SLOW-NEXT: vpshufb %ymm29, %ymm19, %ymm17
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm17, %ymm15 {%k7}
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm17, %ymm14 {%k7}
; AVX512BW-SLOW-NEXT: vpshufb %ymm30, %ymm16, %ymm16
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm16, %ymm14 {%k5}
-; AVX512BW-SLOW-NEXT: vpshufb %ymm21, %ymm19, %ymm16
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm16, %ymm15 {%k5}
+; AVX512BW-SLOW-NEXT: vpshufb %ymm28, %ymm19, %ymm16
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm16, %ymm14 {%k6}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm16, %ymm15 {%k6}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm15, %zmm14
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = <u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm10, %xmm16
-; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm12, %xmm17
+; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm8, %xmm16
+; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm10, %xmm17
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,0,0,1]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
-; AVX512BW-SLOW-NEXT: vpermw %ymm10, %ymm12, %ymm10
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm10, %zmm10
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
+; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
+; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm10, %ymm8
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm8, %zmm8
; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm16 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm7, %xmm17
-; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm8, %xmm18
+; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm9, %xmm18
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX512BW-SLOW-NEXT: vprold $16, %xmm7, %xmm7
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm7, %zmm7
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,0,0,1,4,4,4,5]
; AVX512BW-SLOW-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k3
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm7, %zmm10 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm10, %ymm7
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm9[2,1,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero,xmm8[4],zero,xmm8[5],zero,xmm8[6],zero,xmm8[7],zero
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm7, %zmm8 {%k3}
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm8, %ymm7
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[2,1,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero,xmm9[4],zero,xmm9[5],zero,xmm9[6],zero,xmm9[7],zero
; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [6,5,0,7,6,5,0,7,6,5,0,7,1,1,1,1]
-; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm17, %ymm7 {%k5}
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm13[2,1,2,3]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm17, %ymm8
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm8, %ymm7 {%k6}
-; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero,xmm9[4],zero,xmm9[5],zero,xmm9[6],zero,xmm9[7],zero
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
-; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm9, %ymm10 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm9, %ymm8
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm8, %ymm10 {%k2}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm7
-; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm2, %xmm8
-; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm3, %xmm10
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm17, %ymm7 {%k5}
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm13[2,1,2,3]
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm17, %ymm9
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm9, %ymm7 {%k6}
+; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm9 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero,xmm11[4],zero,xmm11[5],zero,xmm11[6],zero,xmm11[7],zero
+; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
+; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm11, %ymm8 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm11, %ymm9
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm9, %ymm8 {%k2}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm0, %xmm8
+; AVX512BW-SLOW-NEXT: vpshufb %xmm15, %xmm2, %xmm9
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm12, %ymm2
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
-; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm0, %xmm3
-; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm1, %xmm8
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512BW-SLOW-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm2 {%k3}
-; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm0
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[2,1,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm17, %ymm0 {%k5}
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm6[2,1,2,3]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm17, %ymm1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm1, %ymm0 {%k6}
-; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm9, %ymm2 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm9, %ymm1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm1, %ymm2 {%k2}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm10, %ymm0
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm1, %xmm2
+; AVX512BW-SLOW-NEXT: vpshufb %xmm16, %xmm3, %xmm8
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512BW-SLOW-NEXT: vprold $16, %xmm1, %xmm1
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5]
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k3}
+; AVX512BW-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm4[2,1,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm17, %ymm1 {%k5}
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[2,1,2,3]
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm17, %ymm2
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm2, %ymm1 {%k6}
+; AVX512BW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm11, %ymm0 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm11, %ymm2
+; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm2, %ymm0 {%k2}
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm7, 192(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm14, 128(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm20, 320(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm11, 256(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm12, 256(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm5, 64(%rax)
; AVX512BW-SLOW-NEXT: vzeroupper
; AVX512BW-SLOW-NEXT: retq
;
; AVX512BW-FAST-LABEL: store_i8_stride6_vf64:
; AVX512BW-FAST: # %bb.0:
-; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FAST-NEXT: vmovdqa (%rcx), %ymm14
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512BW-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm0, %ymm2
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm3
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[16],ymm2[16],ymm3[17],ymm2[17],ymm3[18],ymm2[18],ymm3[19],ymm2[19],ymm3[20],ymm2[20],ymm3[21],ymm2[21],ymm3[22],ymm2[22],ymm3[23],ymm2[23]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15],ymm1[24],ymm0[24],ymm1[25],ymm0[25],ymm1[26],ymm0[26],ymm1[27],ymm0[27],ymm1[28],ymm0[28],ymm1[29],ymm0[29],ymm1[30],ymm0[30],ymm1[31],ymm0[31]
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-FAST-NEXT: vpermw %ymm3, %ymm8, %ymm3
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm3, %ymm5
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rdx), %ymm4
+; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm4, %ymm6
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[16],ymm5[16],ymm6[17],ymm5[17],ymm6[18],ymm5[18],ymm6[19],ymm5[19],ymm6[20],ymm5[20],ymm6[21],ymm5[21],ymm6[22],ymm5[22],ymm6[23],ymm5[23]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15],ymm4[24],ymm3[24],ymm4[25],ymm3[25],ymm4[26],ymm3[26],ymm4[27],ymm3[27],ymm4[28],ymm3[28],ymm4[29],ymm3[29],ymm4[30],ymm3[30],ymm4[31],ymm3[31]
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512BW-FAST-NEXT: vpermw %ymm6, %ymm12, %ymm6
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm6
+; AVX512BW-FAST-NEXT: movl $613566756, %eax # imm = 0x24924924
+; AVX512BW-FAST-NEXT: kmovd %eax, %k3
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm2, %zmm6 {%k3}
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm6, %ymm10
+; AVX512BW-FAST-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = <10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm5
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512BW-FAST-NEXT: movw $-28087, %ax # imm = 0x9249
+; AVX512BW-FAST-NEXT: kmovd %eax, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm5, %ymm10 {%k1}
+; AVX512BW-FAST-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
+; AVX512BW-FAST-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = <u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
+; AVX512BW-FAST-NEXT: vpshufb %ymm16, %ymm5, %ymm11
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512BW-FAST-NEXT: movl $-2113396606, %eax # imm = 0x82082082
+; AVX512BW-FAST-NEXT: kmovd %eax, %k2
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm11, %ymm10 {%k2}
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = <6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm17, %ymm2, %ymm11
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512BW-FAST-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512BW-FAST-NEXT: kmovd %eax, %k4
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm11, %ymm6 {%k4}
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = <u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm18, %ymm5, %ymm11
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512BW-FAST-NEXT: movl $136348168, %eax # imm = 0x8208208
+; AVX512BW-FAST-NEXT: kmovd %eax, %k5
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm11, %ymm6 {%k5}
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
+; AVX512BW-FAST-NEXT: vmovdqa (%rsi), %ymm10
+; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm10, %ymm13
+; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %ymm11
+; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm11, %ymm7
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm13[0],ymm7[1],ymm13[1],ymm7[2],ymm13[2],ymm7[3],ymm13[3],ymm7[4],ymm13[4],ymm7[5],ymm13[5],ymm7[6],ymm13[6],ymm7[7],ymm13[7],ymm7[16],ymm13[16],ymm7[17],ymm13[17],ymm7[18],ymm13[18],ymm7[19],ymm13[19],ymm7[20],ymm13[20],ymm7[21],ymm13[21],ymm7[22],ymm13[22],ymm7[23],ymm13[23]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm13 = ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15],ymm11[24],ymm10[24],ymm11[25],ymm10[25],ymm11[26],ymm10[26],ymm11[27],ymm10[27],ymm11[28],ymm10[28],ymm11[29],ymm10[29],ymm11[30],ymm10[30],ymm11[31],ymm10[31]
+; AVX512BW-FAST-NEXT: vpermw %ymm13, %ymm8, %ymm8
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512BW-FAST-NEXT: vmovdqa (%rcx), %ymm13
+; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm13, %ymm8
; AVX512BW-FAST-NEXT: vmovdqa (%rdx), %ymm15
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm15[0],ymm14[0],ymm15[1],ymm14[1],ymm15[2],ymm14[2],ymm15[3],ymm14[3],ymm15[4],ymm14[4],ymm15[5],ymm14[5],ymm15[6],ymm14[6],ymm15[7],ymm14[7],ymm15[16],ymm14[16],ymm15[17],ymm14[17],ymm15[18],ymm14[18],ymm15[19],ymm14[19],ymm15[20],ymm14[20],ymm15[21],ymm14[21],ymm15[22],ymm14[22],ymm15[23],ymm14[23]
-; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [9,8,11,10,9,8,11,10,9,8,11,10,13,12,15,14]
-; AVX512BW-FAST-NEXT: vpermw %ymm0, %ymm12, %ymm2
-; AVX512BW-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512BW-FAST-NEXT: vmovdqa 32(%rcx), %xmm7
-; AVX512BW-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-FAST-NEXT: vmovdqa 32(%rdx), %xmm8
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
-; AVX512BW-FAST-NEXT: vpermw %ymm3, %ymm20, %ymm3
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm4
-; AVX512BW-FAST-NEXT: vmovdqa64 (%rsi), %ymm17
-; AVX512BW-FAST-NEXT: vmovdqa64 (%rdi), %ymm18
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm18[0],ymm17[0],ymm18[1],ymm17[1],ymm18[2],ymm17[2],ymm18[3],ymm17[3],ymm18[4],ymm17[4],ymm18[5],ymm17[5],ymm18[6],ymm17[6],ymm18[7],ymm17[7],ymm18[16],ymm17[16],ymm18[17],ymm17[17],ymm18[18],ymm17[18],ymm18[19],ymm17[19],ymm18[20],ymm17[20],ymm18[21],ymm17[21],ymm18[22],ymm17[22],ymm18[23],ymm17[23]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [8,11,10,9,8,11,10,9,8,11,10,9,12,13,14,13]
-; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm21, %ymm5
-; AVX512BW-FAST-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512BW-FAST-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512BW-FAST-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7]
-; AVX512BW-FAST-NEXT: vpermw %ymm6, %ymm22, %ymm6
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512BW-FAST-NEXT: movl $613566756, %r10d # imm = 0x24924924
-; AVX512BW-FAST-NEXT: kmovd %r10d, %k3
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm4, %zmm5 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm5, %ymm23
-; AVX512BW-FAST-NEXT: vmovdqa64 (%r8), %ymm16
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = <2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm24, %ymm16, %ymm4
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512BW-FAST-NEXT: movw $18724, %r10w # imm = 0x4924
-; AVX512BW-FAST-NEXT: kmovd %r10d, %k1
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm4, %ymm23 {%k1}
-; AVX512BW-FAST-NEXT: vmovdqa64 (%r9), %ymm19
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = <u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm19, %ymm4
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512BW-FAST-NEXT: movl $545392672, %r10d # imm = 0x20820820
-; AVX512BW-FAST-NEXT: kmovd %r10d, %k2
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm4, %ymm23 {%k2}
-; AVX512BW-FAST-NEXT: vmovdqa (%r8), %xmm4
-; AVX512BW-FAST-NEXT: vmovdqa 32(%r8), %xmm11
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
-; AVX512BW-FAST-NEXT: movw $-28087, %r10w # imm = 0x9249
-; AVX512BW-FAST-NEXT: kmovd %r10d, %k4
-; AVX512BW-FAST-NEXT: vpermw %ymm6, %ymm26, %ymm5 {%k4}
-; AVX512BW-FAST-NEXT: vmovdqa (%r9), %xmm6
-; AVX512BW-FAST-NEXT: vmovdqa 32(%r9), %xmm13
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm27 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FAST-NEXT: vpermw %ymm27, %ymm26, %ymm27
-; AVX512BW-FAST-NEXT: movl $-2113396606, %r10d # imm = 0x82082082
-; AVX512BW-FAST-NEXT: kmovd %r10d, %k7
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm27, %ymm5 {%k7}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm5, %zmm5
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%rcx), %ymm23
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%rdx), %ymm27
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm23[0],ymm27[1],ymm23[1],ymm27[2],ymm23[2],ymm27[3],ymm23[3],ymm27[4],ymm23[4],ymm27[5],ymm23[5],ymm27[6],ymm23[6],ymm27[7],ymm23[7],ymm27[16],ymm23[16],ymm27[17],ymm23[17],ymm27[18],ymm23[18],ymm27[19],ymm23[19],ymm27[20],ymm23[20],ymm27[21],ymm23[21],ymm27[22],ymm23[22],ymm27[23],ymm23[23]
-; AVX512BW-FAST-NEXT: vpermw %ymm28, %ymm12, %ymm12
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm28 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
-; AVX512BW-FAST-NEXT: vpermw %ymm28, %ymm20, %ymm20
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm20, %zmm12
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%rsi), %ymm20
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%rdi), %ymm28
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm20[0],ymm28[1],ymm20[1],ymm28[2],ymm20[2],ymm28[3],ymm20[3],ymm28[4],ymm20[4],ymm28[5],ymm20[5],ymm28[6],ymm20[6],ymm28[7],ymm20[7],ymm28[16],ymm20[16],ymm28[17],ymm20[17],ymm28[18],ymm20[18],ymm28[19],ymm20[19],ymm28[20],ymm20[20],ymm28[21],ymm20[21],ymm28[22],ymm20[22],ymm28[23],ymm20[23]
-; AVX512BW-FAST-NEXT: vpermw %ymm29, %ymm21, %ymm21
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm29 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
-; AVX512BW-FAST-NEXT: vpermw %ymm29, %ymm22, %ymm22
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm22, %zmm21
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm12, %zmm21 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm21, %ymm12
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%r8), %ymm22
-; AVX512BW-FAST-NEXT: vpshufb %ymm24, %ymm22, %ymm24
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm24, %ymm12 {%k1}
-; AVX512BW-FAST-NEXT: vmovdqa64 32(%r9), %ymm24
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm24, %ymm25
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm25, %ymm12 {%k2}
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm11[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FAST-NEXT: vpermw %ymm25, %ymm26, %ymm21 {%k4}
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FAST-NEXT: vpermw %ymm25, %ymm26, %ymm25
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm25, %ymm21 {%k7}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm21, %zmm12
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = <8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm21, %ymm20, %ymm25
-; AVX512BW-FAST-NEXT: vpshufb %ymm21, %ymm28, %ymm26
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm26[0],ymm25[0],ymm26[1],ymm25[1],ymm26[2],ymm25[2],ymm26[3],ymm25[3],ymm26[4],ymm25[4],ymm26[5],ymm25[5],ymm26[6],ymm25[6],ymm26[7],ymm25[7],ymm26[16],ymm25[16],ymm26[17],ymm25[17],ymm26[18],ymm25[18],ymm26[19],ymm25[19],ymm26[20],ymm25[20],ymm26[21],ymm25[21],ymm26[22],ymm25[22],ymm26[23],ymm25[23]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,2,2,3]
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm20 = ymm28[8],ymm20[8],ymm28[9],ymm20[9],ymm28[10],ymm20[10],ymm28[11],ymm20[11],ymm28[12],ymm20[12],ymm28[13],ymm20[13],ymm28[14],ymm20[14],ymm28[15],ymm20[15],ymm28[24],ymm20[24],ymm28[25],ymm20[25],ymm28[26],ymm20[26],ymm28[27],ymm20[27],ymm28[28],ymm20[28],ymm28[29],ymm20[29],ymm28[30],ymm20[30],ymm28[31],ymm20[31]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-FAST-NEXT: vpermw %ymm20, %ymm26, %ymm20
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm25, %zmm20
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u,5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm23, %ymm28
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm27, %ymm29
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm23 = ymm27[8],ymm23[8],ymm27[9],ymm23[9],ymm27[10],ymm23[10],ymm27[11],ymm23[11],ymm27[12],ymm23[12],ymm27[13],ymm23[13],ymm27[14],ymm23[14],ymm27[15],ymm23[15],ymm27[24],ymm23[24],ymm27[25],ymm23[25],ymm27[26],ymm23[26],ymm27[27],ymm23[27],ymm27[28],ymm23[28],ymm27[29],ymm23[29],ymm27[30],ymm23[30],ymm27[31],ymm23[31]
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512BW-FAST-NEXT: vpermw %ymm23, %ymm27, %ymm23
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm28, %zmm23
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm20, %zmm23 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm23, %ymm20
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = <10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm28, %ymm22, %ymm29
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm29[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm29, %ymm20 {%k4}
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm29 = <u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
-; AVX512BW-FAST-NEXT: vpshufb %ymm29, %ymm24, %ymm30
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm30, %ymm20 {%k7}
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = <6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm30, %ymm22, %ymm22
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
-; AVX512BW-FAST-NEXT: movw $9362, %cx # imm = 0x2492
-; AVX512BW-FAST-NEXT: kmovd %ecx, %k5
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm22, %ymm23 {%k5}
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = <u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %ymm22, %ymm24, %ymm24
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
-; AVX512BW-FAST-NEXT: movl $136348168, %ecx # imm = 0x8208208
-; AVX512BW-FAST-NEXT: kmovd %ecx, %k6
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm24, %ymm23 {%k6}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm23, %zmm20
-; AVX512BW-FAST-NEXT: vpshufb %ymm21, %ymm17, %ymm23
-; AVX512BW-FAST-NEXT: vpshufb %ymm21, %ymm18, %ymm21
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm21 = ymm21[0],ymm23[0],ymm21[1],ymm23[1],ymm21[2],ymm23[2],ymm21[3],ymm23[3],ymm21[4],ymm23[4],ymm21[5],ymm23[5],ymm21[6],ymm23[6],ymm21[7],ymm23[7],ymm21[16],ymm23[16],ymm21[17],ymm23[17],ymm21[18],ymm23[18],ymm21[19],ymm23[19],ymm21[20],ymm23[20],ymm21[21],ymm23[21],ymm21[22],ymm23[22],ymm21[23],ymm23[23]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,2,2,3]
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm17 = ymm18[8],ymm17[8],ymm18[9],ymm17[9],ymm18[10],ymm17[10],ymm18[11],ymm17[11],ymm18[12],ymm17[12],ymm18[13],ymm17[13],ymm18[14],ymm17[14],ymm18[15],ymm17[15],ymm18[24],ymm17[24],ymm18[25],ymm17[25],ymm18[26],ymm17[26],ymm18[27],ymm17[27],ymm18[28],ymm17[28],ymm18[29],ymm17[29],ymm18[30],ymm17[30],ymm18[31],ymm17[31]
-; AVX512BW-FAST-NEXT: vpermw %ymm17, %ymm26, %ymm17
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm21, %zmm17
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm14, %ymm18
-; AVX512BW-FAST-NEXT: vpshufb %ymm25, %ymm15, %ymm21
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm21[0],ymm18[0],ymm21[1],ymm18[1],ymm21[2],ymm18[2],ymm21[3],ymm18[3],ymm21[4],ymm18[4],ymm21[5],ymm18[5],ymm21[6],ymm18[6],ymm21[7],ymm18[7],ymm21[16],ymm18[16],ymm21[17],ymm18[17],ymm21[18],ymm18[18],ymm21[19],ymm18[19],ymm21[20],ymm18[20],ymm21[21],ymm18[21],ymm21[22],ymm18[22],ymm21[23],ymm18[23]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm18[2,2,2,3]
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm14 = ymm15[8],ymm14[8],ymm15[9],ymm14[9],ymm15[10],ymm14[10],ymm15[11],ymm14[11],ymm15[12],ymm14[12],ymm15[13],ymm14[13],ymm15[14],ymm14[14],ymm15[15],ymm14[15],ymm15[24],ymm14[24],ymm15[25],ymm14[25],ymm15[26],ymm14[26],ymm15[27],ymm14[27],ymm15[28],ymm14[28],ymm15[29],ymm14[29],ymm15[30],ymm14[30],ymm15[31],ymm14[31]
-; AVX512BW-FAST-NEXT: vpermw %ymm14, %ymm27, %ymm14
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm18, %zmm14
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm17, %zmm14 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm14, %ymm15
-; AVX512BW-FAST-NEXT: vpshufb %ymm28, %ymm16, %ymm17
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm17, %ymm15 {%k4}
-; AVX512BW-FAST-NEXT: vpshufb %ymm29, %ymm19, %ymm17
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm17, %ymm15 {%k7}
-; AVX512BW-FAST-NEXT: vpshufb %ymm30, %ymm16, %ymm16
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu16 %ymm16, %ymm14 {%k5}
-; AVX512BW-FAST-NEXT: vpshufb %ymm22, %ymm19, %ymm16
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm16, %ymm14 {%k6}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
-; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
-; AVX512BW-FAST-NEXT: vpshufb %xmm15, %xmm7, %xmm16
-; AVX512BW-FAST-NEXT: vpshufb %xmm15, %xmm8, %xmm17
+; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm9
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm15[8],ymm13[8],ymm15[9],ymm13[9],ymm15[10],ymm13[10],ymm15[11],ymm13[11],ymm15[12],ymm13[12],ymm15[13],ymm13[13],ymm15[14],ymm13[14],ymm15[15],ymm13[15],ymm15[24],ymm13[24],ymm15[25],ymm13[25],ymm15[26],ymm13[26],ymm15[27],ymm13[27],ymm15[28],ymm13[28],ymm15[29],ymm13[29],ymm15[30],ymm13[30],ymm15[31],ymm13[31]
+; AVX512BW-FAST-NEXT: vpermw %ymm9, %ymm12, %ymm9
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm12
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm7, %zmm12 {%k3}
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm12, %ymm7
+; AVX512BW-FAST-NEXT: vmovdqa (%r8), %ymm8
+; AVX512BW-FAST-NEXT: vpshufb %ymm14, %ymm8, %ymm9
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm9, %ymm7 {%k1}
+; AVX512BW-FAST-NEXT: vmovdqa (%r9), %ymm9
+; AVX512BW-FAST-NEXT: vpshufb %ymm16, %ymm9, %ymm14
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm14, %ymm7 {%k2}
+; AVX512BW-FAST-NEXT: vpshufb %ymm17, %ymm8, %ymm14
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm14, %ymm12 {%k4}
+; AVX512BW-FAST-NEXT: vpshufb %ymm18, %ymm9, %ymm14
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm14, %ymm12 {%k5}
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm12, %zmm7
+; AVX512BW-FAST-NEXT: vmovdqa64 (%rcx), %xmm21
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm22 = <5,8,7,6,9,u,u,10,u,u,u,u,u,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %xmm22, %xmm12, %xmm16
+; AVX512BW-FAST-NEXT: vmovdqa 32(%rdx), %xmm14
+; AVX512BW-FAST-NEXT: vpshufb %xmm22, %xmm14, %xmm17
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm17[0],xmm16[0],xmm17[1],xmm16[1],xmm17[2],xmm16[2],xmm17[3],xmm16[3],xmm17[4],xmm16[4],xmm17[5],xmm16[5],xmm17[6],xmm16[6],xmm17[7],xmm16[7]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,0,0,1]
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
-; AVX512BW-FAST-NEXT: vpermw %ymm7, %ymm8, %ymm7
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm7, %zmm7
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm16 = <u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u>
-; AVX512BW-FAST-NEXT: vpshufb %xmm16, %xmm9, %xmm17
-; AVX512BW-FAST-NEXT: vpshufb %xmm16, %xmm10, %xmm18
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm17[0,0,0,1]
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3],xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
+; AVX512BW-FAST-NEXT: vpermw %ymm17, %ymm23, %ymm17
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm17, %zmm17
+; AVX512BW-FAST-NEXT: vmovdqa64 32(%rsi), %xmm16
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm24 = <u,u,u,u,u,u,u,u,8,7,6,9,u,u,10,u>
+; AVX512BW-FAST-NEXT: vpshufb %xmm24, %xmm16, %xmm19
+; AVX512BW-FAST-NEXT: vmovdqa64 32(%rdi), %xmm18
+; AVX512BW-FAST-NEXT: vpshufb %xmm24, %xmm18, %xmm20
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm19[0,0,0,1]
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
+; AVX512BW-FAST-NEXT: vpermw %ymm20, %ymm25, %ymm20
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm20, %zmm20
+; AVX512BW-FAST-NEXT: movl $-1840700270, %eax # imm = 0x92492492
+; AVX512BW-FAST-NEXT: kmovd %eax, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm17, %zmm20 {%k1}
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm20, %ymm26
+; AVX512BW-FAST-NEXT: vmovdqa64 32(%r8), %xmm17
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm27 = <8,u,9,u,u,u,u,u,u,u,5,u,6,u,7,u>
+; AVX512BW-FAST-NEXT: vpshufb %xmm27, %xmm17, %xmm19
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [6,5,0,7,6,5,0,7,6,5,0,7,1,1,1,1]
+; AVX512BW-FAST-NEXT: vpermw %ymm19, %ymm28, %ymm26 {%k4}
+; AVX512BW-FAST-NEXT: vmovdqa64 32(%r9), %xmm19
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm29 = <u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7>
+; AVX512BW-FAST-NEXT: vpshufb %xmm29, %xmm19, %xmm30
+; AVX512BW-FAST-NEXT: vpermw %ymm30, %ymm28, %ymm30
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm30, %ymm26 {%k5}
+; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm30 = xmm17[0],zero,xmm17[1],zero,xmm17[2],zero,xmm17[3],zero,xmm17[4],zero,xmm17[5],zero,xmm17[6],zero,xmm17[7],zero
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm31 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
+; AVX512BW-FAST-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512BW-FAST-NEXT: kmovd %eax, %k6
+; AVX512BW-FAST-NEXT: vpermw %ymm30, %ymm31, %ymm20 {%k6}
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm30 = xmm19[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FAST-NEXT: vpermw %ymm30, %ymm31, %ymm30
+; AVX512BW-FAST-NEXT: movl $545392672, %eax # imm = 0x20820820
+; AVX512BW-FAST-NEXT: kmovd %eax, %k7
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm30, %ymm20 {%k7}
+; AVX512BW-FAST-NEXT: vmovdqa64 (%rdx), %xmm30
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm20, %zmm20
+; AVX512BW-FAST-NEXT: vpshufb %xmm22, %xmm21, %xmm26
+; AVX512BW-FAST-NEXT: vpshufb %xmm22, %xmm30, %xmm22
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm22[0],xmm26[0],xmm22[1],xmm26[1],xmm22[2],xmm26[2],xmm22[3],xmm26[3],xmm22[4],xmm26[4],xmm22[5],xmm26[5],xmm22[6],xmm26[6],xmm22[7],xmm26[7]
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm30[0],xmm21[0],xmm30[1],xmm21[1],xmm30[2],xmm21[2],xmm30[3],xmm21[3],xmm30[4],xmm21[4],xmm30[5],xmm21[5],xmm30[6],xmm21[6],xmm30[7],xmm21[7]
+; AVX512BW-FAST-NEXT: vpermw %ymm26, %ymm23, %ymm23
+; AVX512BW-FAST-NEXT: vmovdqa64 (%rsi), %xmm26
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm22[0,0,0,1]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm23, %zmm23
+; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512BW-FAST-NEXT: vpshufb %xmm24, %xmm26, %xmm22
+; AVX512BW-FAST-NEXT: vpshufb %xmm24, %xmm0, %xmm24
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm24[8],xmm22[8],xmm24[9],xmm22[9],xmm24[10],xmm22[10],xmm24[11],xmm22[11],xmm24[12],xmm22[12],xmm24[13],xmm22[13],xmm24[14],xmm22[14],xmm24[15],xmm22[15]
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm24 = xmm0[0],xmm26[0],xmm0[1],xmm26[1],xmm0[2],xmm26[2],xmm0[3],xmm26[3],xmm0[4],xmm26[4],xmm0[5],xmm26[5],xmm0[6],xmm26[6],xmm0[7],xmm26[7]
+; AVX512BW-FAST-NEXT: vpermw %ymm24, %ymm25, %ymm24
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm22[0,0,0,1]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm24, %zmm25
+; AVX512BW-FAST-NEXT: vmovdqa64 (%r8), %xmm22
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm23, %zmm25 {%k1}
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm25, %ymm23
+; AVX512BW-FAST-NEXT: vpshufb %xmm27, %xmm22, %xmm24
+; AVX512BW-FAST-NEXT: vpermw %ymm24, %ymm28, %ymm23 {%k4}
+; AVX512BW-FAST-NEXT: vmovdqa64 (%r9), %xmm24
+; AVX512BW-FAST-NEXT: vpshufb %xmm29, %xmm24, %xmm27
+; AVX512BW-FAST-NEXT: vpermw %ymm27, %ymm28, %ymm27
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm27, %ymm23 {%k5}
+; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm27 = xmm22[0],zero,xmm22[1],zero,xmm22[2],zero,xmm22[3],zero,xmm22[4],zero,xmm22[5],zero,xmm22[6],zero,xmm22[7],zero
+; AVX512BW-FAST-NEXT: vpermw %ymm27, %ymm31, %ymm25 {%k6}
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm27 = xmm24[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FAST-NEXT: vpermw %ymm27, %ymm31, %ymm27
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm27, %ymm25 {%k7}
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm23
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm13 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[16],ymm13[16],ymm15[17],ymm13[17],ymm15[18],ymm13[18],ymm15[19],ymm13[19],ymm15[20],ymm13[20],ymm15[21],ymm13[21],ymm15[22],ymm13[22],ymm15[23],ymm13[23]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm30[8],xmm21[8],xmm30[9],xmm21[9],xmm30[10],xmm21[10],xmm30[11],xmm21[11],xmm30[12],xmm21[12],xmm30[13],xmm21[13],xmm30[14],xmm21[14],xmm30[15],xmm21[15]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm13
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[16],ymm10[16],ymm11[17],ymm10[17],ymm11[18],ymm10[18],ymm11[19],ymm10[19],ymm11[20],ymm10[20],ymm11[21],ymm10[21],ymm11[22],ymm10[22],ymm11[23],ymm10[23]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm26[8],xmm0[9],xmm26[9],xmm0[10],xmm26[10],xmm0[11],xmm26[11],xmm0[12],xmm26[12],xmm0[13],xmm26[13],xmm0[14],xmm26[14],xmm0[15],xmm26[15]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
+; AVX512BW-FAST-NEXT: vpermw %zmm0, %zmm10, %zmm0
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
+; AVX512BW-FAST-NEXT: vpermw %zmm13, %zmm11, %zmm0 {%k3}
+; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512BW-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512BW-FAST-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[4],mem[4],ymm1[5],mem[5],ymm1[6],mem[6],ymm1[7],mem[7],ymm1[16],mem[16],ymm1[17],mem[17],ymm1[18],mem[18],ymm1[19],mem[19],ymm1[20],mem[20],ymm1[21],mem[21],ymm1[22],mem[22],ymm1[23],mem[23]
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
+; AVX512BW-FAST-NEXT: vpermw %zmm1, %zmm10, %zmm1
+; AVX512BW-FAST-NEXT: vpermw %zmm3, %zmm11, %zmm1 {%k3}
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm0, %ymm3
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm8
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm8, %ymm3 {%k6}
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u,u,2,u,1,u,0,u,3,u,u,u,u,u,4,u,u>
+; AVX512BW-FAST-NEXT: vpshufb %ymm8, %ymm9, %ymm9
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm9, %ymm3 {%k7}
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm22[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
+; AVX512BW-FAST-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 # 2-byte Reload
+; AVX512BW-FAST-NEXT: vpermw %ymm9, %ymm10, %ymm0 {%k1}
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm24[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm9, %ymm10, %ymm9
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm9
-; AVX512BW-FAST-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
-; AVX512BW-FAST-NEXT: kmovd %ecx, %k3
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm7, %zmm9 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm7
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm17 = <8,u,9,u,u,u,u,u,u,u,5,u,6,u,7,u>
-; AVX512BW-FAST-NEXT: vpshufb %xmm17, %xmm11, %xmm18
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [6,5,0,7,6,5,0,7,6,5,0,7,1,1,1,1]
-; AVX512BW-FAST-NEXT: vpermw %ymm18, %ymm19, %ymm7 {%k5}
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm18 = <u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7>
-; AVX512BW-FAST-NEXT: vpshufb %xmm18, %xmm13, %xmm21
-; AVX512BW-FAST-NEXT: vpermw %ymm21, %ymm19, %ymm21
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm21, %ymm7 {%k6}
-; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm11 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero,xmm11[4],zero,xmm11[5],zero,xmm11[6],zero,xmm11[7],zero
-; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
-; AVX512BW-FAST-NEXT: vpermw %ymm11, %ymm21, %ymm9 {%k1}
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FAST-NEXT: vpermw %ymm11, %ymm21, %ymm11
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm11, %ymm9 {%k2}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm7
-; AVX512BW-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm9
-; AVX512BW-FAST-NEXT: vpshufb %xmm15, %xmm1, %xmm11
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3],xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512BW-FAST-NEXT: vpermw %ymm0, %ymm8, %ymm0
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
-; AVX512BW-FAST-NEXT: vpshufb %xmm16, %xmm2, %xmm1
-; AVX512BW-FAST-NEXT: vpshufb %xmm16, %xmm3, %xmm8
-; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm1[8],xmm8[9],xmm1[9],xmm8[10],xmm1[10],xmm8[11],xmm1[11],xmm8[12],xmm1[12],xmm8[13],xmm1[13],xmm8[14],xmm1[14],xmm8[15],xmm1[15]
-; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm9, %ymm0 {%k2}
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512BW-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm1, %ymm3
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm2, %ymm3 {%k6}
+; AVX512BW-FAST-NEXT: vpshufb %ymm8, %ymm5, %ymm2
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512BW-FAST-NEXT: vmovdqu8 %ymm2, %ymm3 {%k7}
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm17[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm10, %ymm1 {%k1}
+; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm19[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm10, %ymm2
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512BW-FAST-NEXT: vmovdqu16 %zmm0, %zmm1 {%k3}
-; AVX512BW-FAST-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512BW-FAST-NEXT: vpshufb %xmm17, %xmm4, %xmm2
-; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm19, %ymm0 {%k5}
-; AVX512BW-FAST-NEXT: vpshufb %xmm18, %xmm6, %xmm2
-; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm19, %ymm2
-; AVX512BW-FAST-NEXT: vmovdqu8 %ymm2, %ymm0 {%k6}
-; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
-; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm21, %ymm1 {%k1}
-; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm21, %ymm2
; AVX512BW-FAST-NEXT: vmovdqu8 %ymm2, %ymm1 {%k2}
-; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm7, 192(%rax)
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm14, 128(%rax)
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm20, 320(%rax)
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm12, 256(%rax)
-; AVX512BW-FAST-NEXT: vmovdqa64 %zmm5, 64(%rax)
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm23, (%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm20, 192(%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm6, 320(%rax)
; AVX512BW-FAST-NEXT: vzeroupper
; AVX512BW-FAST-NEXT: retq
%in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
index 08e906dea77b..7a2024c6f908 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
@@ -6545,268 +6545,227 @@ define void @store_i8_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-SLOW: # %bb.0:
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-SLOW-NEXT: vmovdqa (%r10), %xmm0
+; AVX512BW-SLOW-NEXT: vmovdqa (%r10), %xmm1
+; AVX512BW-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512BW-SLOW-NEXT: vmovdqa 16(%r10), %xmm11
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r10), %xmm21
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r10), %xmm19
+; AVX512BW-SLOW-NEXT: vmovdqa (%rax), %xmm0
; AVX512BW-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-SLOW-NEXT: vmovdqa 16(%r10), %xmm14
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r10), %xmm19
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r10), %xmm18
-; AVX512BW-SLOW-NEXT: vmovdqa (%rax), %xmm1
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rax), %xmm16
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rax), %xmm20
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rax), %xmm21
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,4,4,4,4,6,5,0,1,4,6,4,6,6,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm2, %ymm4
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm3, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %xmm4
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r9), %xmm22
-; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm6
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r8), %xmm23
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15]
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,4,5,4,5,5,7,0,1,6,5,6,5,7,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm5, %ymm7, %ymm9
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3]
-; AVX512BW-SLOW-NEXT: vpermw %ymm5, %ymm8, %ymm5
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm17
+; AVX512BW-SLOW-NEXT: vmovdqa 16(%rax), %xmm13
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rax), %xmm22
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rax), %xmm20
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %xmm3
+; AVX512BW-SLOW-NEXT: vmovdqa 16(%r9), %xmm15
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %xmm23
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r9), %xmm25
+; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm4
+; AVX512BW-SLOW-NEXT: vmovdqa64 16(%r8), %xmm16
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %xmm24
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%r8), %xmm26
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3,16,17,20,21,20,21,21,23,16,17,22,21,22,21,23,23]
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm7, %zmm2
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3,16,17,20,20,20,20,22,21,16,17,20,22,20,22,22,23]
; AVX512BW-SLOW-NEXT: movl $-2004318072, %eax # imm = 0x88888888
; AVX512BW-SLOW-NEXT: kmovd %eax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm17 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %xmm9
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rsi), %xmm24
-; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm10
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rdi), %xmm25
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm8, %zmm2 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rsi), %xmm28
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rdi), %xmm29
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm6[8],xmm9[9],xmm6[9],xmm9[10],xmm6[10],xmm9[11],xmm6[11],xmm9[12],xmm6[12],xmm9[13],xmm6[13],xmm9[14],xmm6[14],xmm9[15],xmm6[15]
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm11 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm5, %ymm5
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm11 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm0[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm5, %ymm5
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm10, %ymm0
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %xmm11
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rcx), %xmm27
+; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rcx), %xmm31
; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512BW-SLOW-NEXT: vmovdqa64 48(%rdx), %xmm31
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm13 = [4,4,2,3,6,5,6,5,4,6,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm13, %ymm26
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm15, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7,20,20,18,19,22,21,22,21,20,22,18,19,22,23,22,23]
; AVX512BW-SLOW-NEXT: movl $572662306, %eax # imm = 0x22222222
; AVX512BW-SLOW-NEXT: kmovd %eax, %k2
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm5 {%k2}
-; AVX512BW-SLOW-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512BW-SLOW-NEXT: kmovd %eax, %k3
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm17, %zmm5 {%k3}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm21[0],xmm18[0],xmm21[1],xmm18[1],xmm21[2],xmm18[2],xmm21[3],xmm18[3],xmm21[4],xmm18[4],xmm21[5],xmm18[5],xmm21[6],xmm18[6],xmm21[7],xmm18[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm2, %ymm17
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm3, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm17, %ymm7, %ymm26
-; AVX512BW-SLOW-NEXT: vpermw %ymm17, %ymm8, %ymm17
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm17, %zmm26
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm26 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3],xmm25[4],xmm24[4],xmm25[5],xmm24[5],xmm25[6],xmm24[6],xmm25[7],xmm24[7]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm17 = xmm0[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm17 = xmm17[0],zero,zero,zero,xmm17[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm28 = xmm0[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm28 = xmm28[0],zero,zero,zero,xmm28[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm28, %ymm17, %ymm17
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm28 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm28, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm17
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm31[0],xmm27[0],xmm31[1],xmm27[1],xmm31[2],xmm27[2],xmm31[3],xmm27[3],xmm31[4],xmm27[4],xmm31[5],xmm27[5],xmm31[6],xmm27[6],xmm31[7],xmm27[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm13, %ymm28
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm15, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %xmm28
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm17 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %xmm29
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm26, %zmm17 {%k3}
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rsi), %xmm26
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm21[8],xmm18[8],xmm21[9],xmm18[9],xmm21[10],xmm18[10],xmm21[11],xmm18[11],xmm21[12],xmm18[12],xmm21[13],xmm18[13],xmm21[14],xmm18[14],xmm21[15],xmm18[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm2, %ymm18
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm3, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdi), %xmm30
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm18, %ymm7, %ymm21
-; AVX512BW-SLOW-NEXT: vpermw %ymm18, %ymm8, %ymm18
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm21
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rcx), %xmm22
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm21 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15]
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm14, %zmm5 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm26[0],xmm25[0],xmm26[1],xmm25[1],xmm26[2],xmm25[2],xmm26[3],xmm25[3],xmm26[4],xmm25[4],xmm26[5],xmm25[5],xmm26[6],xmm25[6],xmm26[7],xmm25[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm17, %zmm17, %zmm17
+; AVX512BW-SLOW-NEXT: vpermw %zmm17, %zmm7, %zmm17
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm8, %zmm17 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7]
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm18 = xmm0[2,3,2,3]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm23 = xmm0[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm23, %ymm18, %ymm18
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm27 = xmm0[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm27[0],zero,zero,zero,xmm27[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm27, %ymm18, %ymm18
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm23, %ymm23
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm23, %zmm18
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm31[8],xmm27[8],xmm31[9],xmm27[9],xmm31[10],xmm27[10],xmm31[11],xmm27[11],xmm31[12],xmm27[12],xmm31[13],xmm27[13],xmm31[14],xmm27[14],xmm31[15],xmm27[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm13, %ymm24
-; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm15, %ymm23
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm23, %zmm23
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm23, %zmm18 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm21, %zmm18 {%k3}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm21, %ymm2, %ymm23
-; AVX512BW-SLOW-NEXT: vpermw %ymm21, %ymm3, %ymm21
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm21, %zmm21
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm7, %ymm24
-; AVX512BW-SLOW-NEXT: vpermw %ymm23, %ymm8, %ymm23
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm23, %zmm23
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm21, %zmm23 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm30[0],xmm26[0],xmm30[1],xmm26[1],xmm30[2],xmm26[2],xmm30[3],xmm26[3],xmm30[4],xmm26[4],xmm30[5],xmm26[5],xmm30[6],xmm26[6],xmm30[7],xmm26[7]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm24 = xmm21[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm25 = xmm21[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm25[0],zero,zero,zero,xmm25[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm25, %ymm24, %ymm24
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm21 = xmm21[1,1,1,1]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm21, %ymm25, %ymm21
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm21, %zmm21
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm24 = xmm0[0],xmm22[0],xmm0[1],xmm22[1],xmm0[2],xmm22[2],xmm0[3],xmm22[3],xmm0[4],xmm22[4],xmm0[5],xmm22[5],xmm0[6],xmm22[6],xmm0[7],xmm22[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm24, %ymm13, %ymm25
-; AVX512BW-SLOW-NEXT: vpermw %ymm24, %ymm15, %ymm24
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm25
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%r9), %xmm24
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm25, %zmm21 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%r8), %xmm25
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm23, %zmm21 {%k3}
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rsi), %xmm23
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm27, %ymm0
+; AVX512BW-SLOW-NEXT: vmovdqa 48(%rdx), %xmm1
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm0, %zmm18
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm31[0],xmm1[1],xmm31[1],xmm1[2],xmm31[2],xmm1[3],xmm31[3],xmm1[4],xmm31[4],xmm1[5],xmm31[5],xmm1[6],xmm31[6],xmm1[7],xmm31[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm14, %zmm18 {%k2}
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rsi), %xmm30
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm19, %ymm2, %ymm20
-; AVX512BW-SLOW-NEXT: vpermw %ymm19, %ymm3, %ymm19
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm19
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rdi), %xmm27
+; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rcx), %xmm27
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm19, %zmm25
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm20, %zmm20, %zmm19
+; AVX512BW-SLOW-NEXT: vpermw %zmm19, %zmm7, %zmm19
+; AVX512BW-SLOW-NEXT: vpermw %zmm25, %zmm8, %zmm19 {%k1}
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm20, %ymm7, %ymm28
-; AVX512BW-SLOW-NEXT: vpermw %ymm20, %ymm8, %ymm20
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm20, %zmm28
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rcx), %xmm20
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm19, %zmm28 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm30[8],xmm26[8],xmm30[9],xmm26[9],xmm30[10],xmm26[10],xmm30[11],xmm26[11],xmm30[12],xmm26[12],xmm30[13],xmm26[13],xmm30[14],xmm26[14],xmm30[15],xmm26[15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm26 = xmm19[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm25 = xmm20[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm25[0],zero,zero,zero,xmm25[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm26 = xmm20[3,3,3,3]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm26 = xmm26[0],zero,zero,zero,xmm26[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm29 = xmm19[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm29[0],zero,zero,zero,xmm29[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm29, %ymm26, %ymm29
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm26 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm19 = xmm19[1,1,1,1]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm19 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm19, %ymm26, %ymm19
-; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rdx), %xmm26
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm19, %zmm19
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm22[8],xmm0[9],xmm22[9],xmm0[10],xmm22[10],xmm0[11],xmm22[11],xmm0[12],xmm22[12],xmm0[13],xmm22[13],xmm0[14],xmm22[14],xmm0[15],xmm22[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm13, %ymm22
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm15, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm19 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm28, %zmm19 {%k3}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm16[0],xmm14[0],xmm16[1],xmm14[1],xmm16[2],xmm14[2],xmm16[3],xmm14[3],xmm16[4],xmm14[4],xmm16[5],xmm14[5],xmm16[6],xmm14[6],xmm16[7],xmm14[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm2, %ymm22
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm3, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3],xmm25[4],xmm24[4],xmm25[5],xmm24[5],xmm25[6],xmm24[6],xmm25[7],xmm24[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm22, %ymm7, %ymm28
-; AVX512BW-SLOW-NEXT: vpermw %ymm22, %ymm8, %ymm22
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm22, %zmm28
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm28 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm27[0],xmm23[0],xmm27[1],xmm23[1],xmm27[2],xmm23[2],xmm27[3],xmm23[3],xmm27[4],xmm23[4],xmm27[5],xmm23[5],xmm27[6],xmm23[6],xmm27[7],xmm23[7]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm22 = xmm0[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm26, %ymm25, %ymm25
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm26 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm20 = xmm20[1,1,1,1]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm20, %ymm26, %ymm20
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm20
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm31[8],xmm1[9],xmm31[9],xmm1[10],xmm31[10],xmm1[11],xmm31[11],xmm1[12],xmm31[12],xmm1[13],xmm31[13],xmm1[14],xmm31[14],xmm1[15],xmm31[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm14, %zmm20 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3],xmm22[4],xmm21[4],xmm22[5],xmm21[5],xmm22[6],xmm21[6],xmm22[7],xmm21[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm24[0],xmm23[0],xmm24[1],xmm23[1],xmm24[2],xmm23[2],xmm24[3],xmm23[3],xmm24[4],xmm23[4],xmm24[5],xmm23[5],xmm24[6],xmm23[6],xmm24[7],xmm23[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm25, %zmm25, %zmm25
+; AVX512BW-SLOW-NEXT: vpermw %zmm25, %zmm7, %zmm25
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm8, %zmm25 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm0[0],xmm30[0],xmm0[1],xmm30[1],xmm0[2],xmm30[2],xmm0[3],xmm30[3],xmm0[4],xmm30[4],xmm0[5],xmm30[5],xmm0[6],xmm30[6],xmm0[7],xmm30[7]
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm26 = xmm1[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm26 = xmm26[0],zero,zero,zero,xmm26[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm28 = xmm1[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm28 = xmm28[0],zero,zero,zero,xmm28[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm28, %ymm26, %ymm26
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm28 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm1, %ymm28, %ymm1
+; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdx), %xmm31
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm1, %zmm26
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm31[0],xmm27[0],xmm31[1],xmm27[1],xmm31[2],xmm27[2],xmm31[3],xmm27[3],xmm31[4],xmm27[4],xmm31[5],xmm27[5],xmm31[6],xmm27[6],xmm31[7],xmm27[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm14, %zmm26 {%k2}
+; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rsi), %xmm28
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
+; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rdi), %xmm29
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm24[8],xmm23[8],xmm24[9],xmm23[9],xmm24[10],xmm23[10],xmm24[11],xmm23[11],xmm24[12],xmm23[12],xmm24[13],xmm23[13],xmm24[14],xmm23[14],xmm24[15],xmm23[15]
+; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rcx), %xmm23
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm21, %zmm21, %zmm21
+; AVX512BW-SLOW-NEXT: vpermw %zmm21, %zmm7, %zmm21
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm8, %zmm21 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm30[8],xmm0[9],xmm30[9],xmm0[10],xmm30[10],xmm0[11],xmm30[11],xmm0[12],xmm30[12],xmm0[13],xmm30[13],xmm0[14],xmm30[14],xmm0[15],xmm30[15]
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm22 = xmm0[3,3,3,3]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm22[0],zero,zero,zero,xmm22[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm29 = xmm0[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm29[0],zero,zero,zero,xmm29[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm29, %ymm22, %ymm22
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm22, %ymm1, %ymm1
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm29, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm22
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm26[0],xmm20[0],xmm26[1],xmm20[1],xmm26[2],xmm20[2],xmm26[3],xmm20[3],xmm26[4],xmm20[4],xmm26[5],xmm20[5],xmm26[6],xmm20[6],xmm26[7],xmm20[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm13, %ymm29
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm15, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm22 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm28, %zmm22 {%k3}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm16[8],xmm14[8],xmm16[9],xmm14[9],xmm16[10],xmm14[10],xmm16[11],xmm14[11],xmm16[12],xmm14[12],xmm16[13],xmm14[13],xmm16[14],xmm14[14],xmm16[15],xmm14[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm2, %ymm14
-; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm3, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm14, %ymm7, %ymm16
-; AVX512BW-SLOW-NEXT: vpermw %ymm14, %ymm8, %ymm14
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm14, %zmm14
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm14 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm27[8],xmm23[8],xmm27[9],xmm23[9],xmm27[10],xmm23[10],xmm27[11],xmm23[11],xmm27[12],xmm23[12],xmm27[13],xmm23[13],xmm27[14],xmm23[14],xmm27[15],xmm23[15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm16 = xmm0[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm16 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm23 = xmm0[3,3,3,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm23, %ymm16, %ymm16
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm22, %ymm0
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm22
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm31[8],xmm27[8],xmm31[9],xmm27[9],xmm31[10],xmm27[10],xmm31[11],xmm27[11],xmm31[12],xmm27[12],xmm31[13],xmm27[13],xmm31[14],xmm27[14],xmm31[15],xmm27[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm14, %zmm22 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm7, %zmm24
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm8, %zmm24 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7]
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm27 = xmm0[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm27[0],zero,zero,zero,xmm27[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm27, %ymm1, %ymm1
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm27, %ymm0
+; AVX512BW-SLOW-NEXT: vmovdqa64 16(%rdx), %xmm30
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm27
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm30[0],xmm23[0],xmm30[1],xmm23[1],xmm30[2],xmm23[2],xmm30[3],xmm23[3],xmm30[4],xmm23[4],xmm30[5],xmm23[5],xmm30[6],xmm23[6],xmm30[7],xmm23[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm14, %zmm27 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
+; AVX512BW-SLOW-NEXT: vpermw %zmm1, %zmm7, %zmm1
+; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm8, %zmm1 {%k1}
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15]
+; AVX512BW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload
+; AVX512BW-SLOW-NEXT: # xmm11 = xmm11[0],mem[0],xmm11[1],mem[1],xmm11[2],mem[2],xmm11[3],mem[3],xmm11[4],mem[4],xmm11[5],mem[5],xmm11[6],mem[6],xmm11[7],mem[7]
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm3, %zmm3, %zmm3
+; AVX512BW-SLOW-NEXT: vpermw %zmm3, %zmm7, %zmm3
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm11, %zmm11, %zmm4
+; AVX512BW-SLOW-NEXT: vpermw %zmm4, %zmm8, %zmm3 {%k1}
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm0[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm4, %ymm4
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti32x4 $1, %xmm0, %ymm23, %ymm0
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm26[8],xmm20[8],xmm26[9],xmm20[9],xmm26[10],xmm20[10],xmm26[11],xmm20[11],xmm26[12],xmm20[12],xmm26[13],xmm20[13],xmm26[14],xmm20[14],xmm26[15],xmm20[15]
-; AVX512BW-SLOW-NEXT: vpermw %ymm16, %ymm13, %ymm20
-; AVX512BW-SLOW-NEXT: vpermw %ymm16, %ymm15, %ymm16
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm16, %zmm16
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm16, %zmm0 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm14, %zmm0 {%k3}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512BW-SLOW-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm2, %ymm2
-; AVX512BW-SLOW-NEXT: vpermw %ymm1, %ymm3, %ymm1
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3],xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm7, %ymm3
-; AVX512BW-SLOW-NEXT: vpermw %ymm2, %ymm8, %ymm2
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm1, %zmm2 {%k1}
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm1[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm7, %ymm0
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm30[8],xmm23[8],xmm30[9],xmm23[9],xmm30[10],xmm23[10],xmm30[11],xmm23[11],xmm30[12],xmm23[12],xmm30[13],xmm23[13],xmm30[14],xmm23[14],xmm30[15],xmm23[15]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm4, %zmm4
+; AVX512BW-SLOW-NEXT: vpermw %zmm4, %zmm14, %zmm0 {%k2}
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3],xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[2,3,2,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[3,3,3,3]
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6
+; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
+; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,1,1]
; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
-; AVX512BW-SLOW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm4, %ymm1
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512BW-SLOW-NEXT: vpermw %ymm3, %ymm13, %ymm4
-; AVX512BW-SLOW-NEXT: vpermw %ymm3, %ymm15, %ymm3
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm3, %zmm1 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm2, %zmm1 {%k3}
+; AVX512BW-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm7, %ymm4
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
+; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
+; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm6, %zmm6, %zmm6
+; AVX512BW-SLOW-NEXT: vpermw %zmm6, %zmm14, %zmm4 {%k2}
+; AVX512BW-SLOW-NEXT: movw $-21846, %ax # imm = 0xAAAA
+; AVX512BW-SLOW-NEXT: kmovd %eax, %k1
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm2, %zmm5 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm17, %zmm18 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm19, %zmm20 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm25, %zmm26 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm21, %zmm22 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm24, %zmm27 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa32 %zmm3, %zmm4 {%k1}
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm22, 128(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm19, 320(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm21, 256(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm18, 448(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm17, 384(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm27, 128(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm22, 320(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm26, 256(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm20, 448(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm18, 384(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm5, 64(%rax)
; AVX512BW-SLOW-NEXT: vzeroupper
; AVX512BW-SLOW-NEXT: retq
More information about the llvm-commits
mailing list