[llvm] 06bf4bf - [X86] combineINSERT_SUBVECTOR - enable shuffle combining of concat(extractsub(shuffle),extractsub(shuffle)) patterns (#209510)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 15 01:22:06 PDT 2026
Author: Simon Pilgrim
Date: 2026-07-15T09:22:01+01:00
New Revision: 06bf4bfff830207da473c06f2d90f93d777b51bb
URL: https://github.com/llvm/llvm-project/commit/06bf4bfff830207da473c06f2d90f93d777b51bb
DIFF: https://github.com/llvm/llvm-project/commit/06bf4bfff830207da473c06f2d90f93d777b51bb.diff
LOG: [X86] combineINSERT_SUBVECTOR - enable shuffle combining of concat(extractsub(shuffle),extractsub(shuffle)) patterns (#209510)
Reuse the peekThroughBitcastsAndExtracts helper that we already use for insertsub(shuffle,extractsub(shuffle)) patterns
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d3fff2d4b54e2..d127eea066dde 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -61527,6 +61527,13 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
}
}
+ auto peekThroughBitcastsAndExtracts = [](SDValue V) {
+ while (V.getOpcode() == ISD::BITCAST ||
+ V.getOpcode() == ISD::EXTRACT_SUBVECTOR)
+ V = V.getOperand(0);
+ return V;
+ };
+
// Match concat_vector style patterns.
SmallVector<SDValue, 2> SubVectorOps;
if (collectConcatOps(N, SubVectorOps, DAG)) {
@@ -61546,8 +61553,9 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
SubVectorOps[0], DAG.getVectorIdxConstant(0, dl));
// Attempt to recursively combine to a shuffle.
- if (all_of(SubVectorOps, [](SDValue SubOp) {
- return isTargetShuffle(peekThroughBitcasts(SubOp).getOpcode());
+ if (all_of(SubVectorOps, [&](SDValue SubOp) {
+ SubOp = peekThroughBitcastsAndExtracts(SubOp);
+ return isTargetShuffle(SubOp.getOpcode());
})) {
SDValue Op(N, 0);
if (SDValue Res = combineX86ShufflesRecursively(Op, DAG, Subtarget))
@@ -61601,13 +61609,6 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
}
}
- auto peekThroughBitcastsAndExtracts = [](SDValue V) {
- while (V.getOpcode() == ISD::BITCAST ||
- V.getOpcode() == ISD::EXTRACT_SUBVECTOR)
- V = V.getOperand(0);
- return V;
- };
-
// Attempt to recursively combine to a shuffle.
if (isTargetShuffle(peekThroughBitcasts(Vec).getOpcode()) &&
isTargetShuffle(peekThroughBitcastsAndExtracts(SubVec).getOpcode())) {
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 333fffe20681c..f2505ad876e42 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
@@ -4118,186 +4118,186 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-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
; AVX512-FCP-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
; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [4,5,4,5,4,5,4,5,21,22,21,22,21,22,23,23]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [4,5,4,5,4,5,4,5,21,22,21,22,21,22,23,23]
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm1
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [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]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm0
; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm4, %ymm2
; AVX512-FCP-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]
; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm28
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [2,1,2,3,11,11,11,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm18, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,2,3,11,11,11,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm0
; AVX512-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-FCP-NEXT: kmovw %eax, %k1
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm5
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm2
-; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm4, %ymm2
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [8,21,10,11,20,13,14,23]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [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]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm27
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm21, %zmm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [8,21,10,11,20,13,14,23]
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm6, %zmm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm16, %ymm0
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm29
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,2,0,3,10,0,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm1
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm5
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm3
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm28
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,0,3,10,0,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm2
+; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm5
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm1
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm30
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm22
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm2
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm31
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm20
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm9
-; AVX512-FCP-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
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm1 = 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
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm9[4],ymm5[4],ymm9[5],ymm5[5],ymm9[6],ymm5[6],ymm9[7],ymm5[7],ymm9[12],ymm5[12],ymm9[13],ymm5[13],ymm9[14],ymm5[14],ymm9[15],ymm5[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm16
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm0
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm1
+; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm3, %zmm22
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm3
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,0,0,0,8,8,0,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm19
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm0 = 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
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm2 = 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
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm14[4],ymm4[4],ymm14[5],ymm4[5],ymm14[6],ymm4[6],ymm14[7],ymm4[7],ymm14[12],ymm4[12],ymm14[13],ymm4[13],ymm14[14],ymm4[14],ymm14[15],ymm4[15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm18, %zmm0
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm9, %ymm4
; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm6
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm18, %zmm1
-; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm14
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm4
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm0
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm21, %zmm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm26
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm12, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm10
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm10
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm5
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = 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]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm7[4],ymm9[4],ymm7[5],ymm9[5],ymm7[6],ymm9[6],ymm7[7],ymm9[7],ymm7[12],ymm9[12],ymm7[13],ymm9[13],ymm7[14],ymm9[14],ymm7[15],ymm9[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm5
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm10
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm10, %ymm0
+; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm5, %zmm6
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm16, %ymm5
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm5[0,1,2,3],zmm6[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm15, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm11
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm11
; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm24
; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm25
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm16, %zmm1
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm3
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm11, %xmm13
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [8,8,0,9,0,1,0,1]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm12, %xmm8
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm0[4],xmm8[5],xmm0[5],xmm8[6],xmm0[6],xmm8[7],xmm0[7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [8,8,0,9,0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3]
; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm18
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm0
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm0
; AVX512-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-FCP-NEXT: kmovw %eax, %k2
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k2}
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [16,9,10,17,12,13,18,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm13
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm13[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm23
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm21, %zmm23
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm5
; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm4
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm15[0],xmm1[0],xmm15[1],xmm1[1],xmm15[2],xmm1[2],xmm15[3],xmm1[3]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm15
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm13 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm16, %zmm13
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm2
; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm1
; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm3
; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm19, %zmm3
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT: vpermt2d %ymm6, %ymm17, %ymm21
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm19
-; AVX512-FCP-NEXT: vmovdqa32 %zmm15, %zmm3 {%k2}
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm20, %zmm3
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm6, %ymm16, %ymm0
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm0[0,1,2,3],zmm23[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa32 %zmm13, %zmm3 {%k2}
; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm0
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512-FCP-NEXT: vpermi2d %ymm6, %ymm3, %ymm17
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm23, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm17, %zmm17
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
+; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm3, %zmm21
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512-FCP-NEXT: vpermt2d %ymm6, %ymm16, %ymm3
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm3[0,1,2,3],zmm21[0,1,2,3]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[8],ymm9[8],ymm7[9],ymm9[9],ymm7[10],ymm9[10],ymm7[11],ymm9[11]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [1,1,1,1,10,10,10,11]
; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm1
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm3
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm9[0],ymm3[0],ymm9[1],ymm3[1],ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[8],ymm3[8],ymm9[9],ymm3[9],ymm9[10],ymm3[10],ymm9[11],ymm3[11]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm14[0],ymm3[0],ymm14[1],ymm3[1],ymm14[2],ymm3[2],ymm14[3],ymm3[3],ymm14[8],ymm3[8],ymm14[9],ymm3[9],ymm14[10],ymm3[10],ymm14[11],ymm3[11]
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm3
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm5, %zmm2
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT: vmovdqa %ymm3, %ymm2
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm5, %ymm2
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [8,9,20,11,12,21,14,15]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm14[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm7, %ymm3
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm2[0,1,2,3]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm9 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm0
; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm12
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0],ymm12[0],ymm8[1],ymm12[1],ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[8],ymm12[8],ymm8[9],ymm12[9],ymm8[10],ymm12[10],ymm8[11],ymm12[11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm12
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm6, %zmm11
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [0,0,0,1,0,10,10,0]
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm10, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm13
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm9 = 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]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm13
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm12
; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm25, %xmm8
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm8 = ymm8[0],mem[0],ymm8[1],mem[1],ymm8[2],mem[2],ymm8[3],mem[3],ymm8[8],mem[8],ymm8[9],mem[9],ymm8[10],mem[10],ymm8[11],mem[11]
-; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm8
-; AVX512-FCP-NEXT: vmovdqa32 %zmm11, %zmm8 {%k1}
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512-FCP-NEXT: vpermi2d %ymm1, %ymm8, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm25, %xmm9
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm9 = ymm9[0],mem[0],ymm9[1],mem[1],ymm9[2],mem[2],ymm9[3],mem[3],ymm9[8],mem[8],ymm9[9],mem[9],ymm9[10],mem[10],ymm9[11],mem[11]
+; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm9
+; AVX512-FCP-NEXT: vmovdqa32 %zmm12, %zmm9 {%k1}
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm4
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm5
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm7, %ymm9
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm9[0,1,2,3],zmm5[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm4
; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm4
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm2
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm10, %zmm2
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm4 & (zmm2 ^ zmm1))
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
@@ -4305,16 +4305,16 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm3))
; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm0 & (zmm10 ^ zmm26))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 ^ (zmm0 & (zmm11 ^ zmm26))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 128(%rax)
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
; AVX512-FCP-NEXT: # zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ mem))
; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 ^ (zmm0 & (zmm22 ^ zmm19))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm20 = zmm20 ^ (zmm0 & (zmm20 ^ zmm17))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm20, (%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 ^ (zmm0 & (zmm22 ^ zmm20))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm0 & (zmm19 ^ zmm16))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm19, (%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 192(%rax)
; AVX512-FCP-NEXT: addq $104, %rsp
; AVX512-FCP-NEXT: vzeroupper
@@ -4539,186 +4539,186 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-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-FCP-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-FCP-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-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [4,5,4,5,4,5,4,5,21,22,21,22,21,22,23,23]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [4,5,4,5,4,5,4,5,21,22,21,22,21,22,23,23]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [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-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm4, %ymm2
; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vmovdqa64 %ymm3, %ymm28
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [2,1,2,3,11,11,11,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm18, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm0
; AVX512DQ-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FCP-NEXT: kmovw %eax, %k1
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm5
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm4, %ymm2
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [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-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm27
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm21, %zmm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm6, %zmm2
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm16, %ymm0
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm29
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,2,0,3,10,0,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm3
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm28
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,0,3,10,0,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm5
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm1
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm2
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm30
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm22
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm2
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm31
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm20
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm9
-; AVX512DQ-FCP-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
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm1 = 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
-; AVX512DQ-FCP-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-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm9[4],ymm5[4],ymm9[5],ymm5[5],ymm9[6],ymm5[6],ymm9[7],ymm5[7],ymm9[12],ymm5[12],ymm9[13],ymm5[13],ymm9[14],ymm5[14],ymm9[15],ymm5[15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm16
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm8, %ymm1
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm3, %zmm22
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,0,0,0,8,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm19
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm0 = 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-FCP-NEXT: vpsrldq {{.*#+}} ymm2 = 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-FCP-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-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm14[4],ymm4[4],ymm14[5],ymm4[5],ymm14[6],ymm4[6],ymm14[7],ymm4[7],ymm14[12],ymm4[12],ymm14[13],ymm4[13],ymm14[14],ymm4[14],ymm14[15],ymm4[15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm18, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm9, %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm6
-; AVX512DQ-FCP-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]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm18, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm14
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm0, %ymm1, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm0
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm21, %zmm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm10
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm5
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = 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]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm7[4],ymm9[4],ymm7[5],ymm9[5],ymm7[6],ymm9[6],ymm7[7],ymm9[7],ymm7[12],ymm9[12],ymm7[13],ymm9[13],ymm7[14],ymm9[14],ymm7[15],ymm9[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm10, %ymm0
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm5, %zmm6
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm16, %ymm5
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm5[0,1,2,3],zmm6[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm15, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm11
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm11
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm24
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm25
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm16, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm3
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm11, %xmm13
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [8,8,0,9,0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm12, %xmm8
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm0[4],xmm8[5],xmm0[5],xmm8[6],xmm0[6],xmm8[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [8,8,0,9,0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3]
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm18
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm0
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm0
; AVX512DQ-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FCP-NEXT: kmovw %eax, %k2
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm13
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm13[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm23
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm21, %zmm23
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm5
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm4
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm15[0],xmm1[0],xmm15[1],xmm1[1],xmm15[2],xmm1[2],xmm15[3],xmm1[3]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm15
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm13 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm16, %zmm13
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1
; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm3
; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm19, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm17, %ymm21
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm19
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm15, %zmm3 {%k2}
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm20, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm16, %ymm0
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm0[0,1,2,3],zmm23[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm13, %zmm3 {%k2}
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm0
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm6, %ymm3, %ymm17
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm23, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm17, %zmm17
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm3, %zmm21
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm16, %ymm3
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm3[0,1,2,3],zmm21[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[8],ymm9[8],ymm7[9],ymm9[9],ymm7[10],ymm9[10],ymm7[11],ymm9[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [1,1,1,1,10,10,10,11]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm1
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm3
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm9[0],ymm3[0],ymm9[1],ymm3[1],ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[8],ymm3[8],ymm9[9],ymm3[9],ymm9[10],ymm3[10],ymm9[11],ymm3[11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm14[0],ymm3[0],ymm14[1],ymm3[1],ymm14[2],ymm3[2],ymm14[3],ymm3[3],ymm14[8],ymm3[8],ymm14[9],ymm3[9],ymm14[10],ymm3[10],ymm14[11],ymm3[11]
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm4, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm5, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm5, %ymm2
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm14[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm7, %ymm3
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm2[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm9 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm0
; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,0,1,0,10,10,0]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm12
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0],ymm12[0],ymm8[1],ymm12[1],ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[8],ymm12[8],ymm8[9],ymm12[9],ymm8[10],ymm12[10],ymm8[11],ymm12[11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm12
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm6, %zmm11
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm10, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm9 = 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]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm13
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm12
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm25, %xmm8
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm8 = ymm8[0],mem[0],ymm8[1],mem[1],ymm8[2],mem[2],ymm8[3],mem[3],ymm8[8],mem[8],ymm8[9],mem[9],ymm8[10],mem[10],ymm8[11],mem[11]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm11, %zmm8 {%k1}
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm1, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm25, %xmm9
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm9 = ymm9[0],mem[0],ymm9[1],mem[1],ymm9[2],mem[2],ymm9[3],mem[3],ymm9[8],mem[8],ymm9[9],mem[9],ymm9[10],mem[10],ymm9[11],mem[11]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm4, %zmm9
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm12, %zmm9 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm4
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm7, %ymm9
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm9[0,1,2,3],zmm5[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm4
; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm4
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm2
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm10, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm4 & (zmm2 ^ zmm1))
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
@@ -4726,16 +4726,16 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm3))
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm0 & (zmm10 ^ zmm26))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 ^ (zmm0 & (zmm11 ^ zmm26))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 128(%rax)
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT: # zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ mem))
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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-FCP-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 ^ (zmm0 & (zmm22 ^ zmm19))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm20 = zmm20 ^ (zmm0 & (zmm20 ^ zmm17))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, (%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 ^ (zmm0 & (zmm22 ^ zmm20))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm0 & (zmm19 ^ zmm16))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, (%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 192(%rax)
; AVX512DQ-FCP-NEXT: addq $104, %rsp
; AVX512DQ-FCP-NEXT: vzeroupper
@@ -8680,22 +8680,22 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-FCP-NEXT: kmovw %eax, %k1
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [8,21,10,11,20,13,14,23]
; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm4
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [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]
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm3
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm3
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm3
-; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm20, %ymm3
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [8,21,10,11,20,13,14,23]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm1
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm23, %ymm2
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm3[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm2
; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [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]
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm2
@@ -8706,8 +8706,8 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm2
; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm1
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,0,8,8,0,9]
; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm26
@@ -8728,18 +8728,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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]
; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm2
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1
; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm4
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm3
+; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm20, %zmm1
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm20, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm3
-; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm23, %zmm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm23, %ymm2
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm2
; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm2
; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm2
; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -8767,16 +8767,16 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-FCP-NEXT: kmovw %eax, %k2
; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k2}
-; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm3
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm5
; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm28, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm27, %zmm3
; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm27, %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm28, %ymm1
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm3[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm4
; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -8797,19 +8797,19 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm1
; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k2}
-; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm5
; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm28, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm27, %zmm3
; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm27, %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm28, %ymm1
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm3[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [0,0,0,0,8,8,0,9]
; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm3
; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm1
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,0,2,1,4,5,6,7]
; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm3
; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -8832,18 +8832,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-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]
; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm22, %zmm3
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
-; AVX512-FCP-NEXT: vmovdqa %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm5
; AVX512-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm4
+; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm20, %zmm1
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm20, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm4
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm23, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm23, %ymm3
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm3
; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm1
; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm3
; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -8856,21 +8856,21 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm21, %zmm1
; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm10
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm3
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm9
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm7[4],ymm10[4],ymm7[5],ymm10[5],ymm7[6],ymm10[6],ymm7[7],ymm10[7],ymm7[12],ymm10[12],ymm7[13],ymm10[13],ymm7[14],ymm10[14],ymm7[15],ymm10[15]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm0
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm12
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm20
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm12, %ymm1
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm20, %zmm24
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm1
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm11
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm11, %ymm1
+; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm20
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm23, %ymm0
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm0[0,1,2,3],zmm20[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm12, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm1
; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm1
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
@@ -8888,18 +8888,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm5, %xmm3
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm18
+; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm16
; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm17
; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm0
; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k2}
-; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm4
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm0
+; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm27, %zmm1
; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm16
-; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm27, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm21
+; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm28, %ymm0
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm1[0,1,2,3]
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm6
; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm5
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
@@ -8909,110 +8909,110 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm3
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm4
; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm7
; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm2
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm8
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm8 {%k2}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
-; AVX512-FCP-NEXT: vpermi2d %ymm3, %ymm8, %ymm27
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm8
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm29, %zmm2
; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm15
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm15[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm31, %zmm9
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm27, %zmm22
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm8
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm7[0],ymm10[0],ymm7[1],ymm10[1],ymm7[2],ymm10[2],ymm7[3],ymm10[3],ymm7[8],ymm10[8],ymm7[9],ymm10[9],ymm7[10],ymm10[10],ymm7[11],ymm10[11]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm7
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm15[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm31, %zmm8
+; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm2 {%k2}
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm3
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermi2d %zmm7, %zmm2, %zmm27
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512-FCP-NEXT: vpermt2d %ymm7, %ymm28, %ymm2
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm2[0,1,2,3],zmm27[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[8],ymm10[8],ymm9[9],ymm10[9],ymm9[10],ymm10[10],ymm9[11],ymm10[11]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [1,1,1,1,10,10,10,11]
; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm13[0],ymm14[0],ymm13[1],ymm14[1],ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[8],ymm14[8],ymm13[9],ymm14[9],ymm13[10],ymm14[10],ymm13[11],ymm14[11]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm13[0],ymm14[0],ymm13[1],ymm14[1],ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[8],ymm14[8],ymm13[9],ymm14[9],ymm13[10],ymm14[10],ymm13[11],ymm14[11]
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm29 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm4
-; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm4 {%k1}
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT: vmovdqa %ymm4, %ymm1
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [8,9,20,11,12,21,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm4
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm27
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm1, %zmm2
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = 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]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm3
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm7 = ymm7[0],mem[0],ymm7[1],mem[1],ymm7[2],mem[2],ymm7[3],mem[3],ymm7[8],mem[8],ymm7[9],mem[9],ymm7[10],mem[10],ymm7[11],mem[11]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm7
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm7 {%k1}
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %ymm7, %ymm3
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm3
-; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm7
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm28
-; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm7
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm1, %zmm7
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = 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]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm3
-; AVX512-FCP-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm5
+; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm5 {%k1}
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm1
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm4, %zmm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm25, %ymm5
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm5[0,1,2,3],zmm1[0,1,2,3]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,0,0,1,0,10,10,0]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm6 = ymm6[0],mem[0],ymm6[1],mem[1],ymm6[2],mem[2],ymm6[3],mem[3],ymm6[8],mem[8],ymm6[9],mem[9],ymm6[10],mem[10],ymm6[11],mem[11]
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm29, %zmm6
+; AVX512-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}
+; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm5
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm25, %ymm6
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm6[0,1,2,3],zmm5[0,1,2,3]
+; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm1 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm0, %zmm6
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
+; AVX512-FCP-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512-FCP-NEXT: # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm10
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm10 {%k1}
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %ymm10, %ymm3
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm3
-; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm10
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm29, %zmm10
+; AVX512-FCP-NEXT: vmovdqa32 %zmm5, %zmm10 {%k1}
+; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm5
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm5
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm25, %ymm10
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm5[0,1,2,3]
; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
; AVX512-FCP-NEXT: # ymm10 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm0, %zmm1
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512-FCP-NEXT: # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm11
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm11
; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm12
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm11
@@ -9024,32 +9024,32 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: # ymm12 = 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]
; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm29, %zmm12
; AVX512-FCP-NEXT: vmovdqa32 %zmm11, %zmm12 {%k1}
-; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm10
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm10, %xmm5
-; AVX512-FCP-NEXT: vpermi2d %ymm5, %ymm12, %ymm6
-; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm5 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm25, %zmm12
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm6, %zmm5
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm15, %xmm4
-; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm1, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm1 & (zmm4 ^ zmm5))
+; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm10 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermi2d %zmm10, %zmm12, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm10
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm10, %xmm9
+; AVX512-FCP-NEXT: vpermt2d %ymm9, %ymm25, %ymm12
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm12[0,1,2,3],zmm4[0,1,2,3]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm15, %xmm3
+; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm9 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm0, %zmm3
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm0 & (zmm3 ^ zmm4))
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 256(%rax)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ zmm3))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 448(%rax)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm1 & (zmm7 ^ zmm28))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 640(%rax)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm27))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 256(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 448(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm0 & (zmm6 ^ zmm28))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 640(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm0 & (zmm2 ^ zmm27))
; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm0 & (zmm8 ^ zmm22))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm0 & (zmm9 ^ zmm21))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm0 & (zmm7 ^ zmm22))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm0 & (zmm8 ^ zmm21))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm24))
@@ -9532,22 +9532,22 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FCP-NEXT: kmovw %eax, %k1
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [8,21,10,11,20,13,14,23]
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [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-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm3
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm20, %zmm3
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm20 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm3
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm20, %ymm3
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm1
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm3[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [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-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm2
@@ -9558,8 +9558,8 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm1
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,0,8,8,0,9]
; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm26
@@ -9580,18 +9580,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm3
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm20, %zmm1
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm20, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm3
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm23, %zmm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm2
; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -9619,16 +9619,16 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FCP-NEXT: kmovw %eax, %k2
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm3
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm5
; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm28, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm27, %zmm3
; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm27, %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm28, %ymm1
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm3[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm4
; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -9649,19 +9649,19 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm5
; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm28, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm27, %zmm3
; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm27, %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm28, %ymm1
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm3[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [0,0,0,0,8,8,0,9]
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm3
; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm1
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,0,2,1,4,5,6,7]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -9684,18 +9684,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-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]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm22, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm5
; AVX512DQ-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm4
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm20, %zmm1
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm20, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm4
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm23, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm23, %ymm3
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm3
; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm1
; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm3
; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -9708,21 +9708,21 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm21, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm10
; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
; AVX512DQ-FCP-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-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm7[4],ymm10[4],ymm7[5],ymm10[5],ymm7[6],ymm10[6],ymm7[7],ymm10[7],ymm7[12],ymm10[12],ymm7[13],ymm10[13],ymm7[14],ymm10[14],ymm7[15],ymm10[15]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm12
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm20
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm12, %ymm1
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm20, %zmm24
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm11, %ymm1
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm20
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm23, %ymm0
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm0[0,1,2,3],zmm20[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm12
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm12, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm1
; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
@@ -9740,18 +9740,18 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm5, %xmm3
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm18
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm16
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm17
; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm4
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm0
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm27, %zmm1
; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm16
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm27, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm21
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm28, %ymm0
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm1[0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm6
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm5
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm0 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
@@ -9761,110 +9761,110 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm4
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm7
; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm8 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm3, %ymm8, %ymm27
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm8
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm29, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm15
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm3
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm15[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm31, %zmm9
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm27, %zmm22
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm8
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm7[0],ymm10[0],ymm7[1],ymm10[1],ymm7[2],ymm10[2],ymm7[3],ymm10[3],ymm7[8],ymm10[8],ymm7[9],ymm10[9],ymm7[10],ymm10[10],ymm7[11],ymm10[11]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm7
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm15[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm31, %zmm8
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm2 {%k2}
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm3
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm7, %zmm2, %zmm27
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm7, %ymm28, %ymm2
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm2[0,1,2,3],zmm27[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[8],ymm10[8],ymm9[9],ymm10[9],ymm9[10],ymm10[10],ymm9[11],ymm10[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [1,1,1,1,10,10,10,11]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm13[0],ymm14[0],ymm13[1],ymm14[1],ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[8],ymm14[8],ymm13[9],ymm14[9],ymm13[10],ymm14[10],ymm13[11],ymm14[11]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm13[0],ymm14[0],ymm13[1],ymm14[1],ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[8],ymm14[8],ymm13[9],ymm14[9],ymm13[10],ymm14[10],ymm13[11],ymm14[11]
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm29 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm4 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm4, %ymm1
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm4
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm27
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,0,0,1,0,10,10,0]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm1, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = 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-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm7 = ymm7[0],mem[0],ymm7[1],mem[1],ymm7[2],mem[2],ymm7[3],mem[3],ymm7[8],mem[8],ymm7[9],mem[9],ymm7[10],mem[10],ymm7[11],mem[11]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm7 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm7
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm28
-; AVX512DQ-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm7
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm1, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = 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-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm5 {%k1}
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm4, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm25, %ymm5
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm5[0,1,2,3],zmm1[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm6 = ymm6[0],mem[0],ymm6[1],mem[1],ymm6[2],mem[2],ymm6[3],mem[3],ymm6[8],mem[8],ymm6[9],mem[9],ymm6[10],mem[10],ymm6[11],mem[11]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm29, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}
+; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm25, %ymm6
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm6[0,1,2,3],zmm5[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm0, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT: # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm29, %zmm10
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm10 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm10, %ymm3
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm10
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm29, %zmm10
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm5, %zmm10 {%k1}
+; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm5
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm25, %ymm10
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm5[0,1,2,3]
; AVX512DQ-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT: # ymm10 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT: # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm11
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm12
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm11
@@ -9876,32 +9876,32 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: # ymm12 = 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-FCP-NEXT: vpermt2d %zmm10, %zmm29, %zmm12
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm11, %zmm12 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm10
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm10, %xmm5
-; AVX512DQ-FCP-NEXT: vpermi2d %ymm5, %ymm12, %ymm6
-; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm25, %zmm12
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm6, %zmm5
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm15, %xmm4
-; AVX512DQ-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm1, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm1 & (zmm4 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm10 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm10, %zmm12, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm10
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm10, %xmm9
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm9, %ymm25, %ymm12
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm12[0,1,2,3],zmm4[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm15, %xmm3
+; AVX512DQ-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm0, %zmm3
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm0 & (zmm3 ^ zmm4))
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 256(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm1 & (zmm0 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 448(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm1 & (zmm7 ^ zmm28))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 640(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm27))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 256(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm0 & (zmm1 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 448(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm0 & (zmm6 ^ zmm28))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 640(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm0 & (zmm2 ^ zmm27))
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [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-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm0 & (zmm8 ^ zmm22))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm0 & (zmm9 ^ zmm21))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm0 & (zmm7 ^ zmm22))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm0 & (zmm8 ^ zmm21))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 192(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm1 & (zmm2 ^ zmm24))
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 d1701fa880897..97e4525cf6129 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
@@ -6287,7 +6287,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm1
-; AVX512BW-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512BW-NEXT: vmovdqa (%rsi), %xmm15
; AVX512BW-NEXT: vmovdqa 32(%rsi), %xmm3
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} xmm4 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-NEXT: vpshufb %xmm4, %xmm3, %xmm0
@@ -6302,56 +6302,56 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa (%rcx), %xmm13
; AVX512BW-NEXT: vmovdqa 32(%rcx), %xmm8
; AVX512BW-NEXT: vpshufb %xmm0, %xmm8, %xmm7
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm15
-; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm9
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm10
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512BW-NEXT: vprold $16, %xmm10, %xmm10
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm7
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm10, %xmm9
+; AVX512BW-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-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm7
; AVX512BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,0,0,1,4,4,4,5]
; AVX512BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512BW-NEXT: kmovd %r10d, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm7, %zmm2 {%k1}
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
; AVX512BW-NEXT: vmovdqa64 (%r8), %xmm17
; AVX512BW-NEXT: vmovdqa 32(%r8), %xmm7
-; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
-; AVX512BW-NEXT: vmovdqa %ymm2, %ymm12
-; AVX512BW-NEXT: vpermt2w %ymm10, %ymm11, %ymm12
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm10 = xmm7[2,1,2,3]
-; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero,xmm10[4],zero,xmm10[5],zero,xmm10[6],zero,xmm10[7],zero
-; AVX512BW-NEXT: vpermt2w %zmm10, %zmm18, %zmm2
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm12, %zmm2
-; AVX512BW-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm10[2,1,2,3]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm7[2,1,2,3]
+; AVX512BW-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-NEXT: vmovdqa64 %zmm2, %zmm12
+; AVX512BW-NEXT: vpermt2w %zmm9, %zmm11, %zmm12
+; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm9 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
+; AVX512BW-NEXT: vpermt2w %ymm9, %ymm18, %ymm2
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm12[0,1,2,3]
+; AVX512BW-NEXT: vmovdqa 32(%r9), %xmm9
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm9[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm9[2,1,2,3]
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm19[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
; AVX512BW-NEXT: vpermt2w %zmm19, %zmm20, %zmm12
; AVX512BW-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820
; AVX512BW-NEXT: kmovq %r10, %k2
; AVX512BW-NEXT: vmovdqu8 %zmm12, %zmm2 {%k2}
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm14, %xmm12
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm15, %xmm12
; AVX512BW-NEXT: vpshufb %xmm4, %xmm16, %xmm4
; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm12[8],xmm4[9],xmm12[9],xmm4[10],xmm12[10],xmm4[11],xmm12[11],xmm4[12],xmm12[12],xmm4[13],xmm12[13],xmm4[14],xmm12[14],xmm4[15],xmm12[15]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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-NEXT: vpermt2w %zmm4, %zmm6, %zmm12
; AVX512BW-NEXT: vpshufb %xmm0, %xmm13, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm15, %xmm6
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm14, %xmm6
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3],xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512BW-NEXT: vprold $16, %xmm6, %xmm6
; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
; AVX512BW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,0,0,1,4,4,4,5]
; AVX512BW-NEXT: vmovdqu16 %zmm4, %zmm12 {%k1}
-; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = 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-NEXT: vpermi2w %ymm4, %ymm12, %ymm11
; AVX512BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm17[2,1,2,3]
; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = 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-NEXT: vpermt2w %zmm4, %zmm18, %zmm12
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm4
+; AVX512BW-NEXT: vpermi2w %zmm4, %zmm12, %zmm11
+; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = 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-NEXT: vpermt2w %ymm4, %ymm18, %ymm12
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm12[0,1,2,3],zmm11[0,1,2,3]
; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-NEXT: vpshufd {{.*#+}} xmm12 = xmm6[2,1,2,3]
@@ -6361,23 +6361,23 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa (%rcx), %ymm11
; AVX512BW-NEXT: vmovdqa (%rdx), %ymm12
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[4],ymm11[4],ymm12[5],ymm11[5],ymm12[6],ymm11[6],ymm12[7],ymm11[7],ymm12[16],ymm11[16],ymm12[17],ymm11[17],ymm12[18],ymm11[18],ymm12[19],ymm11[19],ymm12[20],ymm11[20],ymm12[21],ymm11[21],ymm12[22],ymm11[22],ymm12[23],ymm11[23]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
; AVX512BW-NEXT: vpermt2w %zmm18, %zmm20, %zmm19
; AVX512BW-NEXT: vmovdqa (%rsi), %ymm13
-; AVX512BW-NEXT: vmovdqa (%rdi), %ymm15
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = 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-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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-NEXT: vmovdqa (%rdi), %ymm14
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23]
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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-NEXT: vpmovsxbw {{.*#+}} zmm21 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
; AVX512BW-NEXT: vpermt2w %zmm18, %zmm21, %zmm16
; AVX512BW-NEXT: movl $613566756, %r10d # imm = 0x24924924
; AVX512BW-NEXT: kmovd %r10d, %k2
; AVX512BW-NEXT: vmovdqu16 %zmm19, %zmm16 {%k2}
; AVX512BW-NEXT: vextracti64x4 $1, %zmm16, %ymm18
-; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512BW-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpshufb %ymm19, %ymm14, %ymm22
+; AVX512BW-NEXT: vpshufb %ymm19, %ymm15, %ymm22
; AVX512BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
; AVX512BW-NEXT: movw $18724, %r10w # imm = 0x4924
; AVX512BW-NEXT: kmovd %r10d, %k3
@@ -6398,11 +6398,11 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa64 32(%rcx), %ymm18
; AVX512BW-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm23[0],ymm18[0],ymm23[1],ymm18[1],ymm23[2],ymm18[2],ymm23[3],ymm18[3],ymm23[4],ymm18[4],ymm23[5],ymm18[5],ymm23[6],ymm18[6],ymm23[7],ymm18[7],ymm23[16],ymm18[16],ymm23[17],ymm18[17],ymm23[18],ymm18[18],ymm23[19],ymm18[19],ymm23[20],ymm18[20],ymm23[21],ymm18[21],ymm23[22],ymm18[22],ymm23[23],ymm18[23]
-; AVX512BW-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-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-NEXT: vpermt2w %zmm25, %zmm20, %zmm8
-; AVX512BW-NEXT: vmovdqa 32(%rsi), %ymm9
+; AVX512BW-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512BW-NEXT: vmovdqa64 32(%rdi), %ymm20
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm9[0],ymm20[1],ymm9[1],ymm20[2],ymm9[2],ymm20[3],ymm9[3],ymm20[4],ymm9[4],ymm20[5],ymm9[5],ymm20[6],ymm9[6],ymm20[7],ymm9[7],ymm20[16],ymm9[16],ymm20[17],ymm9[17],ymm20[18],ymm9[18],ymm20[19],ymm9[19],ymm20[20],ymm9[20],ymm20[21],ymm9[21],ymm20[22],ymm9[22],ymm20[23],ymm9[23]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm10[0],ymm20[1],ymm10[1],ymm20[2],ymm10[2],ymm20[3],ymm10[3],ymm20[4],ymm10[4],ymm20[5],ymm10[5],ymm20[6],ymm10[6],ymm20[7],ymm10[7],ymm20[16],ymm10[16],ymm20[17],ymm10[17],ymm20[18],ymm10[18],ymm20[19],ymm10[19],ymm20[20],ymm10[20],ymm20[21],ymm10[21],ymm20[22],ymm10[22],ymm20[23],ymm10[23]
; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm3[8],xmm5[9],xmm3[9],xmm5[10],xmm3[10],xmm5[11],xmm3[11],xmm5[12],xmm3[12],xmm5[13],xmm3[13],xmm5[14],xmm3[14],xmm5[15],xmm3[15]
; AVX512BW-NEXT: vpermt2w %zmm25, %zmm21, %zmm3
; AVX512BW-NEXT: vmovdqu16 %zmm8, %zmm3 {%k2}
@@ -6416,15 +6416,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
; AVX512BW-NEXT: vmovdqa 32(%r9), %ymm5
; AVX512BW-NEXT: vpshufb %ymm17, %ymm5, %ymm5
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm10[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-NEXT: vpermt2w %zmm5, %zmm24, %zmm7
; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm3 {%k4}
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-NEXT: vpshufb %ymm5, %ymm9, %ymm7
-; AVX512BW-NEXT: vpshufb %ymm5, %ymm20, %ymm10
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[4],ymm7[4],ymm10[5],ymm7[5],ymm10[6],ymm7[6],ymm10[7],ymm7[7],ymm10[16],ymm7[16],ymm10[17],ymm7[17],ymm10[18],ymm7[18],ymm10[19],ymm7[19],ymm10[20],ymm7[20],ymm10[21],ymm7[21],ymm10[22],ymm7[22],ymm10[23],ymm7[23]
+; AVX512BW-NEXT: vpshufb %ymm5, %ymm10, %ymm7
+; AVX512BW-NEXT: vpshufb %ymm5, %ymm20, %ymm9
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[16],ymm7[16],ymm9[17],ymm7[17],ymm9[18],ymm7[18],ymm9[19],ymm7[19],ymm9[20],ymm7[20],ymm9[21],ymm7[21],ymm9[22],ymm7[22],ymm9[23],ymm7[23]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm20[8],ymm9[8],ymm20[9],ymm9[9],ymm20[10],ymm9[10],ymm20[11],ymm9[11],ymm20[12],ymm9[12],ymm20[13],ymm9[13],ymm20[14],ymm9[14],ymm20[15],ymm9[15],ymm20[24],ymm9[24],ymm20[25],ymm9[25],ymm20[26],ymm9[26],ymm20[27],ymm9[27],ymm20[28],ymm9[28],ymm20[29],ymm9[29],ymm20[30],ymm9[30],ymm20[31],ymm9[31]
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm20[8],ymm10[8],ymm20[9],ymm10[9],ymm20[10],ymm10[10],ymm20[11],ymm10[11],ymm20[12],ymm10[12],ymm20[13],ymm10[13],ymm20[14],ymm10[14],ymm20[15],ymm10[15],ymm20[24],ymm10[24],ymm20[25],ymm10[25],ymm20[26],ymm10[26],ymm20[27],ymm10[27],ymm20[28],ymm10[28],ymm20[29],ymm10[29],ymm20[30],ymm10[30],ymm20[31],ymm10[31]
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm10 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512BW-NEXT: vpermw %ymm9, %ymm10, %ymm9
; AVX512BW-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
@@ -6447,10 +6447,10 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: kmovq %rcx, %k3
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm9 {%k3}
; AVX512BW-NEXT: vpshufb %ymm5, %ymm13, %ymm1
-; AVX512BW-NEXT: vpshufb %ymm5, %ymm15, %ymm5
+; AVX512BW-NEXT: vpshufb %ymm5, %ymm14, %ymm5
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm5[0],ymm1[0],ymm5[1],ymm1[1],ymm5[2],ymm1[2],ymm5[3],ymm1[3],ymm5[4],ymm1[4],ymm5[5],ymm1[5],ymm5[6],ymm1[6],ymm5[7],ymm1[7],ymm5[16],ymm1[16],ymm5[17],ymm1[17],ymm5[18],ymm1[18],ymm5[19],ymm1[19],ymm5[20],ymm1[20],ymm5[21],ymm1[21],ymm5[22],ymm1[22],ymm5[23],ymm1[23]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm5 = 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-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31]
; AVX512BW-NEXT: vpermw %ymm5, %ymm10, %ymm5
; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
; AVX512BW-NEXT: vpshufb %ymm0, %ymm11, %ymm5
@@ -6461,7 +6461,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vpermw %ymm5, %ymm18, %ymm5
; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm14[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm15[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
@@ -6482,15 +6482,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm1
; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm2
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm2
; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm14
; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm3
; AVX512BW-FCP-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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} zmm5 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm2, %zmm5, %zmm3
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm2, %zmm4, %zmm3
; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm15
; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm7
; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
@@ -6505,20 +6505,20 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512BW-FCP-NEXT: kmovd %r10d, %k1
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm10 {%k1}
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %xmm16
; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %xmm9
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm18 = [8,9,0,0,0,5,6,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm9, %xmm2
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm19
+; AVX512BW-FCP-NEXT: vpermt2w %zmm2, %zmm3, %zmm19
; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm2 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
-; AVX512BW-FCP-NEXT: vmovdqa64 %ymm10, %ymm18
-; AVX512BW-FCP-NEXT: vpermt2w %ymm2, %ymm3, %ymm18
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm19 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm20 = [8,9,0,0,0,5,6,7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm9, %xmm2
-; AVX512BW-FCP-NEXT: vpermt2w %zmm2, %zmm19, %zmm10
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm18, %zmm2
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm20 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
+; AVX512BW-FCP-NEXT: vpermt2w %ymm2, %ymm20, %ymm10
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm10[0,1,2,3],zmm19[0,1,2,3]
; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm18 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm10, %xmm21
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm10, %xmm21
; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm23 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
; AVX512BW-FCP-NEXT: vpermt2w %zmm21, %zmm23, %zmm22
@@ -6529,21 +6529,21 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm14, %xmm22
; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm21, %zmm5, %zmm22
-; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm5
+; AVX512BW-FCP-NEXT: vpermt2w %zmm21, %zmm4, %zmm22
+; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm4
; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm17, %xmm11
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm5[8],xmm11[9],xmm5[9],xmm11[10],xmm5[10],xmm11[11],xmm5[11],xmm11[12],xmm5[12],xmm11[13],xmm5[13],xmm11[14],xmm5[14],xmm11[15],xmm5[15]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm11[8],xmm4[8],xmm11[9],xmm4[9],xmm11[10],xmm4[10],xmm11[11],xmm4[11],xmm11[12],xmm4[12],xmm11[13],xmm4[13],xmm11[14],xmm4[14],xmm11[15],xmm4[15]
; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm5, %zmm12, %zmm11
+; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm12, %zmm11
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm22, %zmm11 {%k1}
-; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm16[0],zero,xmm16[1],zero,xmm16[2],zero,xmm16[3],zero,xmm16[4],zero,xmm16[5],zero,xmm16[6],zero,xmm16[7],zero
-; AVX512BW-FCP-NEXT: vpermi2w %ymm5, %ymm11, %ymm3
-; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm16, %xmm5
-; AVX512BW-FCP-NEXT: vpermt2w %zmm5, %zmm19, %zmm11
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm5, %xmm11
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm5[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm16, %xmm4
+; AVX512BW-FCP-NEXT: vpermi2w %zmm4, %zmm11, %zmm3
+; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm16[0],zero,xmm16[1],zero,xmm16[2],zero,xmm16[3],zero,xmm16[4],zero,xmm16[5],zero,xmm16[6],zero,xmm16[7],zero
+; AVX512BW-FCP-NEXT: vpermt2w %ymm4, %ymm20, %ymm11
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm11[0,1,2,3],zmm3[0,1,2,3]
+; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm4, %xmm11
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-FCP-NEXT: vpermt2w %zmm11, %zmm23, %zmm12
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm12, %zmm3 {%k2}
; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm11
@@ -6576,8 +6576,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm17, %zmm16
; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512BW-FCP-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm17, %ymm5, %ymm18
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT: vpshufb %ymm17, %ymm4, %ymm18
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm24 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
; AVX512BW-FCP-NEXT: vpermt2w %zmm18, %zmm24, %zmm23
; AVX512BW-FCP-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
@@ -6586,27 +6586,27 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %ymm18
; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm23[0],ymm18[0],ymm23[1],ymm18[1],ymm23[2],ymm18[2],ymm23[3],ymm18[3],ymm23[4],ymm18[4],ymm23[5],ymm18[5],ymm23[6],ymm18[6],ymm23[7],ymm18[7],ymm23[16],ymm18[16],ymm23[17],ymm18[17],ymm23[18],ymm18[18],ymm23[19],ymm18[19],ymm23[20],ymm18[20],ymm23[21],ymm18[21],ymm23[22],ymm18[22],ymm23[23],ymm18[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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-FCP-NEXT: vpermt2w %zmm25, %zmm20, %zmm4
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm25, %zmm20, %zmm5
; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm20
; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm6[0],ymm20[1],ymm6[1],ymm20[2],ymm6[2],ymm20[3],ymm6[3],ymm20[4],ymm6[4],ymm20[5],ymm6[5],ymm20[6],ymm6[6],ymm20[7],ymm6[7],ymm20[16],ymm6[16],ymm20[17],ymm6[17],ymm20[18],ymm6[18],ymm20[19],ymm6[19],ymm20[20],ymm6[20],ymm20[21],ymm6[21],ymm20[22],ymm6[22],ymm20[23],ymm6[23]
; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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-FCP-NEXT: vpermt2w %zmm25, %zmm21, %zmm7
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm7 {%k2}
-; AVX512BW-FCP-NEXT: vextracti64x4 $1, %zmm7, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm7 {%k2}
+; AVX512BW-FCP-NEXT: vextracti64x4 $1, %zmm7, %ymm5
; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm8, %ymm19
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm19[2,2,2,3]
-; AVX512BW-FCP-NEXT: vmovdqu16 %ymm19, %ymm4 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqu16 %ymm19, %ymm5 {%k3}
; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FCP-NEXT: vpermt2w %ymm9, %ymm22, %ymm7
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm7
; AVX512BW-FCP-NEXT: vpshufb %ymm17, %ymm7, %ymm7
; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm10[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FCP-NEXT: vpermt2w %zmm7, %zmm24, %zmm9
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm4 {%k4}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm5 {%k4}
; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm6, %ymm9
; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm20, %ymm10
@@ -6652,12 +6652,12 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm15[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 320(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 256(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
@@ -6668,7 +6668,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW: # %bb.0:
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm1
-; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm15
; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %xmm3
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} xmm4 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm3, %xmm0
@@ -6683,56 +6683,56 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm13
; AVX512DQ-BW-NEXT: vmovdqa 32(%rcx), %xmm8
; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm8, %xmm7
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm15
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm9
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm10
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512DQ-BW-NEXT: vprold $16, %xmm10, %xmm10
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm7
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm10, %xmm9
+; AVX512DQ-BW-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]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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]
+; AVX512DQ-BW-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm7
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,0,0,1,4,4,4,5]
; AVX512DQ-BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512DQ-BW-NEXT: kmovd %r10d, %k1
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm7, %zmm2 {%k1}
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %xmm17
; AVX512DQ-BW-NEXT: vmovdqa 32(%r8), %xmm7
-; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
-; AVX512DQ-BW-NEXT: vmovdqa %ymm2, %ymm12
-; AVX512DQ-BW-NEXT: vpermt2w %ymm10, %ymm11, %ymm12
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm10 = xmm7[2,1,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero,xmm10[4],zero,xmm10[5],zero,xmm10[6],zero,xmm10[7],zero
-; AVX512DQ-BW-NEXT: vpermt2w %zmm10, %zmm18, %zmm2
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm12, %zmm2
-; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm10[2,1,2,3]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm7[2,1,2,3]
+; AVX512DQ-BW-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
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm12
+; AVX512DQ-BW-NEXT: vpermt2w %zmm9, %zmm11, %zmm12
+; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm9 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
+; AVX512DQ-BW-NEXT: vpermt2w %ymm9, %ymm18, %ymm2
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm12[0,1,2,3]
+; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %xmm9
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm9[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm9[2,1,2,3]
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm19[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
; AVX512DQ-BW-NEXT: vpermt2w %zmm19, %zmm20, %zmm12
; AVX512DQ-BW-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820
; AVX512DQ-BW-NEXT: kmovq %r10, %k2
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm12, %zmm2 {%k2}
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm14, %xmm12
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm15, %xmm12
; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm16, %xmm4
; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm12[8],xmm4[9],xmm12[9],xmm4[10],xmm12[10],xmm4[11],xmm12[11],xmm4[12],xmm12[12],xmm4[13],xmm12[13],xmm4[14],xmm12[14],xmm4[15],xmm12[15]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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]
; AVX512DQ-BW-NEXT: vpermt2w %zmm4, %zmm6, %zmm12
; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm13, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm15, %xmm6
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm14, %xmm6
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3],xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512DQ-BW-NEXT: vprold $16, %xmm6, %xmm6
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,0,0,1,4,4,4,5]
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm4, %zmm12 {%k1}
-; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm17[0],zero,xmm17[1],zero,xmm17[2],zero,xmm17[3],zero,xmm17[4],zero,xmm17[5],zero,xmm17[6],zero,xmm17[7],zero
-; AVX512DQ-BW-NEXT: vpermi2w %ymm4, %ymm12, %ymm11
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm17[2,1,2,3]
; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
-; AVX512DQ-BW-NEXT: vpermt2w %zmm4, %zmm18, %zmm12
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm4
+; AVX512DQ-BW-NEXT: vpermi2w %zmm4, %zmm12, %zmm11
+; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm17[0],zero,xmm17[1],zero,xmm17[2],zero,xmm17[3],zero,xmm17[4],zero,xmm17[5],zero,xmm17[6],zero,xmm17[7],zero
+; AVX512DQ-BW-NEXT: vpermt2w %ymm4, %ymm18, %ymm12
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm12[0,1,2,3],zmm11[0,1,2,3]
; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm12 = xmm6[2,1,2,3]
@@ -6742,23 +6742,23 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm11
; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm12
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[4],ymm11[4],ymm12[5],ymm11[5],ymm12[6],ymm11[6],ymm12[7],ymm11[7],ymm12[16],ymm11[16],ymm12[17],ymm11[17],ymm12[18],ymm11[18],ymm12[19],ymm11[19],ymm12[20],ymm11[20],ymm12[21],ymm11[21],ymm12[22],ymm11[22],ymm12[23],ymm11[23]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
; AVX512DQ-BW-NEXT: vpermt2w %zmm18, %zmm20, %zmm19
; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm13
-; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm15
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = 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]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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]
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm14
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm18 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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]
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm21 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
; AVX512DQ-BW-NEXT: vpermt2w %zmm18, %zmm21, %zmm16
; AVX512DQ-BW-NEXT: movl $613566756, %r10d # imm = 0x24924924
; AVX512DQ-BW-NEXT: kmovd %r10d, %k2
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm19, %zmm16 {%k2}
; AVX512DQ-BW-NEXT: vextracti64x4 $1, %zmm16, %ymm18
-; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512DQ-BW-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm19, %ymm14, %ymm22
+; AVX512DQ-BW-NEXT: vpshufb %ymm19, %ymm15, %ymm22
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
; AVX512DQ-BW-NEXT: movw $18724, %r10w # imm = 0x4924
; AVX512DQ-BW-NEXT: kmovd %r10d, %k3
@@ -6779,11 +6779,11 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %ymm18
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm23[0],ymm18[0],ymm23[1],ymm18[1],ymm23[2],ymm18[2],ymm23[3],ymm18[3],ymm23[4],ymm18[4],ymm23[5],ymm18[5],ymm23[6],ymm18[6],ymm23[7],ymm18[7],ymm23[16],ymm18[16],ymm23[17],ymm18[17],ymm23[18],ymm18[18],ymm23[19],ymm18[19],ymm23[20],ymm18[20],ymm23[21],ymm18[21],ymm23[22],ymm18[22],ymm23[23],ymm18[23]
-; AVX512DQ-BW-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]
+; AVX512DQ-BW-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]
; AVX512DQ-BW-NEXT: vpermt2w %zmm25, %zmm20, %zmm8
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %ymm9
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %ymm20
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm9[0],ymm20[1],ymm9[1],ymm20[2],ymm9[2],ymm20[3],ymm9[3],ymm20[4],ymm9[4],ymm20[5],ymm9[5],ymm20[6],ymm9[6],ymm20[7],ymm9[7],ymm20[16],ymm9[16],ymm20[17],ymm9[17],ymm20[18],ymm9[18],ymm20[19],ymm9[19],ymm20[20],ymm9[20],ymm20[21],ymm9[21],ymm20[22],ymm9[22],ymm20[23],ymm9[23]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm10[0],ymm20[1],ymm10[1],ymm20[2],ymm10[2],ymm20[3],ymm10[3],ymm20[4],ymm10[4],ymm20[5],ymm10[5],ymm20[6],ymm10[6],ymm20[7],ymm10[7],ymm20[16],ymm10[16],ymm20[17],ymm10[17],ymm20[18],ymm10[18],ymm20[19],ymm10[19],ymm20[20],ymm10[20],ymm20[21],ymm10[21],ymm20[22],ymm10[22],ymm20[23],ymm10[23]
; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm3[8],xmm5[9],xmm3[9],xmm5[10],xmm3[10],xmm5[11],xmm3[11],xmm5[12],xmm3[12],xmm5[13],xmm3[13],xmm5[14],xmm3[14],xmm5[15],xmm3[15]
; AVX512DQ-BW-NEXT: vpermt2w %zmm25, %zmm21, %zmm3
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm8, %zmm3 {%k2}
@@ -6797,15 +6797,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %ymm5
; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm5, %ymm5
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm10[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQ-BW-NEXT: vpermt2w %zmm5, %zmm24, %zmm7
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm3 {%k4}
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm9, %ymm7
-; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm20, %ymm10
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[4],ymm7[4],ymm10[5],ymm7[5],ymm10[6],ymm7[6],ymm10[7],ymm7[7],ymm10[16],ymm7[16],ymm10[17],ymm7[17],ymm10[18],ymm7[18],ymm10[19],ymm7[19],ymm10[20],ymm7[20],ymm10[21],ymm7[21],ymm10[22],ymm7[22],ymm10[23],ymm7[23]
+; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm10, %ymm7
+; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm20, %ymm9
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[16],ymm7[16],ymm9[17],ymm7[17],ymm9[18],ymm7[18],ymm9[19],ymm7[19],ymm9[20],ymm7[20],ymm9[21],ymm7[21],ymm9[22],ymm7[22],ymm9[23],ymm7[23]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm20[8],ymm9[8],ymm20[9],ymm9[9],ymm20[10],ymm9[10],ymm20[11],ymm9[11],ymm20[12],ymm9[12],ymm20[13],ymm9[13],ymm20[14],ymm9[14],ymm20[15],ymm9[15],ymm20[24],ymm9[24],ymm20[25],ymm9[25],ymm20[26],ymm9[26],ymm20[27],ymm9[27],ymm20[28],ymm9[28],ymm20[29],ymm9[29],ymm20[30],ymm9[30],ymm20[31],ymm9[31]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm20[8],ymm10[8],ymm20[9],ymm10[9],ymm20[10],ymm10[10],ymm20[11],ymm10[11],ymm20[12],ymm10[12],ymm20[13],ymm10[13],ymm20[14],ymm10[14],ymm20[15],ymm10[15],ymm20[24],ymm10[24],ymm20[25],ymm10[25],ymm20[26],ymm10[26],ymm20[27],ymm10[27],ymm20[28],ymm10[28],ymm20[29],ymm10[29],ymm20[30],ymm10[30],ymm20[31],ymm10[31]
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm10 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512DQ-BW-NEXT: vpermw %ymm9, %ymm10, %ymm9
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
@@ -6828,10 +6828,10 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: kmovq %rcx, %k3
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm9 {%k3}
; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm13, %ymm1
-; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm15, %ymm5
+; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm14, %ymm5
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm5[0],ymm1[0],ymm5[1],ymm1[1],ymm5[2],ymm1[2],ymm5[3],ymm1[3],ymm5[4],ymm1[4],ymm5[5],ymm1[5],ymm5[6],ymm1[6],ymm5[7],ymm1[7],ymm5[16],ymm1[16],ymm5[17],ymm1[17],ymm5[18],ymm1[18],ymm5[19],ymm1[19],ymm5[20],ymm1[20],ymm5[21],ymm1[21],ymm5[22],ymm1[22],ymm5[23],ymm1[23]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm5 = 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]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31]
; AVX512DQ-BW-NEXT: vpermw %ymm5, %ymm10, %ymm5
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm11, %ymm5
@@ -6842,7 +6842,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vpermw %ymm5, %ymm18, %ymm5
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm14[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm15[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
@@ -6863,15 +6863,15 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm1
; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm2
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm14
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm3
; AVX512DQ-BW-FCP-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]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm5 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm2, %zmm5, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm2, %zmm4, %zmm3
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm15
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm7
; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
@@ -6886,20 +6886,20 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm10 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %xmm16
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %xmm9
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm18 = [8,9,0,0,0,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm9, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm19
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm2, %zmm3, %zmm19
; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero,xmm9[4],zero,xmm9[5],zero,xmm9[6],zero,xmm9[7],zero
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %ymm10, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm2, %ymm3, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm19 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm20 = [8,9,0,0,0,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm9, %xmm2
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm2, %zmm19, %zmm10
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm18, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm20 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm2, %ymm20, %ymm10
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm10[0,1,2,3],zmm19[0,1,2,3]
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm18 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm10, %xmm21
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm10, %xmm21
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm23 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm21, %zmm23, %zmm22
@@ -6910,21 +6910,21 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm14, %xmm22
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = 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]
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm21, %zmm5, %zmm22
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm5
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm21, %zmm4, %zmm22
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm4
; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm17, %xmm11
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm5[8],xmm11[9],xmm5[9],xmm11[10],xmm5[10],xmm11[11],xmm5[11],xmm11[12],xmm5[12],xmm11[13],xmm5[13],xmm11[14],xmm5[14],xmm11[15],xmm5[15]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm11[8],xmm4[8],xmm11[9],xmm4[9],xmm11[10],xmm4[10],xmm11[11],xmm4[11],xmm11[12],xmm4[12],xmm11[13],xmm4[13],xmm11[14],xmm4[14],xmm11[15],xmm4[15]
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm5, %zmm12, %zmm11
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm12, %zmm11
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm22, %zmm11 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm16[0],zero,xmm16[1],zero,xmm16[2],zero,xmm16[3],zero,xmm16[4],zero,xmm16[5],zero,xmm16[6],zero,xmm16[7],zero
-; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm5, %ymm11, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm16, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm5, %zmm19, %zmm11
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm5, %xmm11
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm5[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm16, %xmm4
+; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm4, %zmm11, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm16[0],zero,xmm16[1],zero,xmm16[2],zero,xmm16[3],zero,xmm16[4],zero,xmm16[5],zero,xmm16[6],zero,xmm16[7],zero
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm4, %ymm20, %ymm11
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm11[0,1,2,3],zmm3[0,1,2,3]
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm4, %xmm11
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm11, %zmm23, %zmm12
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm12, %zmm3 {%k2}
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm11
@@ -6957,8 +6957,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm17, %zmm16
; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512DQ-BW-FCP-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm17, %ymm5, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm17, %ymm4, %ymm18
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm24 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm18, %zmm24, %zmm23
; AVX512DQ-BW-FCP-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
@@ -6967,27 +6967,27 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %ymm18
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm23[0],ymm18[0],ymm23[1],ymm18[1],ymm23[2],ymm18[2],ymm23[3],ymm18[3],ymm23[4],ymm18[4],ymm23[5],ymm18[5],ymm23[6],ymm18[6],ymm23[7],ymm18[7],ymm23[16],ymm18[16],ymm23[17],ymm18[17],ymm23[18],ymm18[18],ymm23[19],ymm18[19],ymm23[20],ymm18[20],ymm23[21],ymm18[21],ymm23[22],ymm18[22],ymm23[23],ymm18[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm25, %zmm20, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm25, %zmm20, %zmm5
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm20
; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm25 = ymm20[0],ymm6[0],ymm20[1],ymm6[1],ymm20[2],ymm6[2],ymm20[3],ymm6[3],ymm20[4],ymm6[4],ymm20[5],ymm6[5],ymm20[6],ymm6[6],ymm20[7],ymm6[7],ymm20[16],ymm6[16],ymm20[17],ymm6[17],ymm20[18],ymm6[18],ymm20[19],ymm6[19],ymm20[20],ymm6[20],ymm20[21],ymm6[21],ymm20[22],ymm6[22],ymm20[23],ymm6[23]
; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm25, %zmm21, %zmm7
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm7 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vextracti64x4 $1, %zmm7, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm7 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vextracti64x4 $1, %zmm7, %ymm5
; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm8, %ymm19
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm19[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm19, %ymm4 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm19, %ymm5 {%k3}
; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm9, %ymm22, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm7
; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm17, %ymm7, %ymm7
; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm10[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm7, %zmm24, %zmm9
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm4 {%k4}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm5 {%k4}
; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm6, %ymm9
; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm20, %ymm10
@@ -7033,12 +7033,12 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm15[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 320(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 256(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
index 7d6ca16313583..45cfdbb0b232f 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512vbmi.ll
@@ -215,11 +215,8 @@ define <64 x i8> @combine_vpermi2var_constant_v64i8_with_mask_commute(<64 x i8>
define <32 x i8> @concat_vpermt2var_v16i8(<32 x i8> %x0, <32 x i8> %x1) {
; CHECK-LABEL: concat_vpermt2var_v16i8:
; CHECK: # %bb.0:
-; CHECK-NEXT: vmovdqa {{.*#+}} xmm2 = [0,17,2,18,4,19,6,21,8,23,10,25,12,27,14,29]
-; CHECK-NEXT: vpermi2b %xmm1, %xmm0, %xmm2
-; CHECK-NEXT: vmovdqa {{.*#+}} xmm3 = [16,49,18,50,20,51,22,53,24,55,26,57,28,59,30,61]
-; CHECK-NEXT: vpermi2b %ymm1, %ymm0, %ymm3
-; CHECK-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm0
+; CHECK-NEXT: vmovdqa {{.*#+}} ymm2 = [0,33,2,34,4,35,6,37,8,39,10,41,12,43,14,45,16,49,18,50,20,51,22,53,24,55,26,57,28,59,30,61]
+; CHECK-NEXT: vpermt2b %ymm1, %ymm2, %ymm0
; CHECK-NEXT: ret{{[l|q]}}
%l0 = shufflevector <32 x i8> %x0, <32 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%l1 = shufflevector <32 x i8> %x1, <32 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
More information about the llvm-commits
mailing list