[llvm] 8bc2d19 - [X86] canonicalizeShuffleWithOp - don't fold VPERMI(BINOP(X,Y)) -> BINOP(VPERMI(X),VPERMI(Y))
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 2 10:38:53 PDT 2024
Author: Simon Pilgrim
Date: 2024-04-02T18:38:37+01:00
New Revision: 8bc2d19c13c3a0aa0bffe82596f75f25bed5004f
URL: https://github.com/llvm/llvm-project/commit/8bc2d19c13c3a0aa0bffe82596f75f25bed5004f
DIFF: https://github.com/llvm/llvm-project/commit/8bc2d19c13c3a0aa0bffe82596f75f25bed5004f.diff
LOG: [X86] canonicalizeShuffleWithOp - don't fold VPERMI(BINOP(X,Y)) -> BINOP(VPERMI(X),VPERMI(Y))
VPERMI (VPERMQ/PD) is nearly always lane-crossing and poorly merges with target shuffles (other than itself).
For now, I've restricted VPERMI to only merge with itself, constants, loads and splats.
We might be able to merge with a few other special cases (AND/ANDNP with constant?), which could help the shuffle-vs-trunc-256.ll AVX512VL regression, but since that now gives similar codegen to the other AVX512 variants, I'd prefer to improve the shuffle lowering for that properly.
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 312e4487a8f17f..a9751e170fe63d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -39961,8 +39961,10 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
const SDLoc &DL) {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
EVT ShuffleVT = N.getValueType();
+ unsigned Opc = N.getOpcode();
- auto IsMergeableWithShuffle = [&DAG](SDValue Op, bool FoldLoad = false) {
+ auto IsMergeableWithShuffle = [Opc, &DAG](SDValue Op, bool FoldShuf = true,
+ bool FoldLoad = false) {
// AllZeros/AllOnes constants are freely shuffled and will peek through
// bitcasts. Other constant build vectors do not peek through bitcasts. Only
// merge with target shuffles if it has one use so shuffle combining is
@@ -39972,8 +39974,9 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()) ||
getTargetConstantFromNode(dyn_cast<LoadSDNode>(Op)) ||
+ (Op.getOpcode() == Opc && Op->hasOneUse()) ||
(Op.getOpcode() == ISD::INSERT_SUBVECTOR && Op->hasOneUse()) ||
- (isTargetShuffle(Op.getOpcode()) && Op->hasOneUse()) ||
+ (FoldShuf && isTargetShuffle(Op.getOpcode()) && Op->hasOneUse()) ||
(FoldLoad && isShuffleFoldableLoad(Op)) ||
DAG.isSplatValue(Op, /*AllowUndefs*/ false);
};
@@ -39984,7 +39987,6 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
(Op.getScalarValueSizeInBits() <= ShuffleVT.getScalarSizeInBits());
};
- unsigned Opc = N.getOpcode();
switch (Opc) {
// Unary and Unary+Permute Shuffles.
case X86ISD::PSHUFB: {
@@ -40010,8 +40012,10 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
if (TLI.isBinOp(SrcOpcode) && IsSafeToMoveShuffle(N0, SrcOpcode)) {
SDValue Op00 = peekThroughOneUseBitcasts(N0.getOperand(0));
SDValue Op01 = peekThroughOneUseBitcasts(N0.getOperand(1));
- if (IsMergeableWithShuffle(Op00, Opc != X86ISD::PSHUFB) ||
- IsMergeableWithShuffle(Op01, Opc != X86ISD::PSHUFB)) {
+ if (IsMergeableWithShuffle(Op00, Opc != X86ISD::VPERMI,
+ Opc != X86ISD::PSHUFB) ||
+ IsMergeableWithShuffle(Op01, Opc != X86ISD::VPERMI,
+ Opc != X86ISD::PSHUFB)) {
SDValue LHS, RHS;
Op00 = DAG.getBitcast(ShuffleVT, Op00);
Op01 = DAG.getBitcast(ShuffleVT, Op01);
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
index 2d7e6f6e58beee..a80f38089c4014 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
@@ -1333,8 +1333,10 @@ define <16 x i8> @negative(<32 x i8> %v, <32 x i8> %w) nounwind {
; AVX512VL-LABEL: negative:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm0[2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]
-; AVX512VL-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,3,2,3]
-; AVX512VL-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm0
+; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
+; AVX512VL-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512VL-NEXT: vpternlogq $206, %ymm1, %ymm0, %ymm2
+; AVX512VL-NEXT: vpermq {{.*#+}} ymm0 = ymm2[0,3,2,3]
; AVX512VL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index 2b539aecc2ad8e..f56c43eb49df2f 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
@@ -2208,15 +2208,13 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
; AVX2-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,ymm3[19,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25],zero,ymm3[23]
-; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25]
-; AVX2-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
; AVX2-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm10 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
; AVX2-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm8
; AVX2-NEXT: vpshufd {{.*#+}} ymm9 = ymm0[0,2,1,1,4,6,5,5]
@@ -2284,15 +2282,13 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
; AVX2-FP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm6
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,ymm3[19,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25],zero,ymm3[23]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vpor %ymm7, %ymm8, %ymm7
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm8 = ymm0[0,2,1,1,4,6,5,5]
@@ -2300,15 +2296,13 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,ymm1[29,26],zero,ymm1[28],zero,ymm1[30],zero,ymm1[28,29],zero,ymm1[31],zero,ymm1[29]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[29,26],zero,ymm3[28],zero,ymm3[26,27,28,29],zero,ymm3[31],zero,ymm3[29,30],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
; AVX2-FP-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm8
; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm9 = ymm0[2,2,3,3,6,6,7,7]
@@ -2375,15 +2369,13 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm6
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,ymm3[19,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25],zero,ymm3[23]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [4,6,5,5,5,5,4,6]
@@ -2391,15 +2383,13 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,ymm1[29,26],zero,ymm1[28],zero,ymm1[30],zero,ymm1[28,29],zero,ymm1[31],zero,ymm1[29]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[29,26],zero,ymm3[28],zero,ymm3[26,27,28,29],zero,ymm3[31],zero,ymm3[29,30],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
; AVX2-FCP-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm8
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [6,6,6,6,7,7,7,7]
@@ -2430,10 +2420,10 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-LABEL: store_i8_stride5_vf32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512-NEXT: vmovdqa (%rdx), %ymm1
+; AVX512-NEXT: vmovdqa (%rcx), %ymm2
; AVX512-NEXT: vmovdqa (%r8), %ymm0
; AVX512-NEXT: vmovdqa (%rdi), %xmm5
; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
@@ -2463,45 +2453,40 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
; AVX512-NEXT: vpermd %zmm6, %zmm8, %zmm6
; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
-; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u]
; AVX512-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u,u],zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u,u,19]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u,u],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u,u,19]
; AVX512-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX512-NEXT: vpternlogq $226, %ymm5, %ymm11, %ymm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21,u],zero,ymm4[20],zero,ymm4[22],zero,ymm4[24,u],zero,ymm4[23],zero,ymm4[25,u]
+; AVX512-NEXT: vpor %ymm5, %ymm9, %ymm5
; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512-NEXT: vpor %ymm9, %ymm10, %ymm9
; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm10 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512-NEXT: vpternlogq $248, %ymm10, %ymm5, %ymm9
-; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm3[21],zero,ymm3[21,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm5
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm9
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm5
; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[12],zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,ymm0[18],zero
; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm0[0,2,1,1,4,6,5,5]
; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,3,2]
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX512-NEXT: vpandn %ymm9, %ymm11, %ymm9
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX512-NEXT: vpandn %ymm9, %ymm10, %ymm9
; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,ymm3[26],zero,ymm3[28],zero,ymm3[30],zero,zero,ymm3[29],zero,ymm3[31],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,ymm3[27],zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm4[26],zero,ymm4[28,u],zero,ymm4[u],zero,ymm4[29],zero,ymm4[31,u],zero,ymm4[30]
+; AVX512-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512-NEXT: vpternlogq $248, %ymm10, %ymm4, %ymm3
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27,u],zero,ymm2[26],zero,ymm2[28],zero,ymm2[30,u],zero,ymm2[29],zero,ymm2[31,u]
+; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512-NEXT: vpternlogq $184, %ymm3, %ymm7, %ymm1
+; AVX512-NEXT: vpternlogq $226, %ymm3, %ymm7, %ymm1
; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
@@ -2513,10 +2498,10 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-FCP-LABEL: store_i8_stride5_vf32:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm4
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm6
@@ -2545,26 +2530,23 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
; AVX512-FCP-NEXT: vpermd %zmm4, %zmm7, %zmm7
; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u]
; AVX512-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u,u],zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u,u,19]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u,u],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u,u,19]
; AVX512-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX512-FCP-NEXT: vpternlogq $226, %ymm5, %ymm10, %ymm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21,u],zero,ymm3[20],zero,ymm3[22],zero,ymm3[24,u],zero,ymm3[23],zero,ymm3[25,u]
+; AVX512-FCP-NEXT: vpor %ymm5, %ymm9, %ymm5
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[21],zero,zero,ymm0[20],zero,ymm0[22],zero,ymm0[24],zero,zero,ymm0[23],zero,ymm0[25],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512-FCP-NEXT: vpor %ymm9, %ymm10, %ymm9
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm10, %ymm5, %ymm9
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm2[21],zero,ymm2[21,20],zero,ymm2[22],zero,ymm2[24],zero,ymm2[22,23],zero,ymm2[25]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512-FCP-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm9
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm5
; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [4,0,5,5,5,5,0,6]
; AVX512-FCP-NEXT: vpermd %ymm4, %ymm8, %ymm8
@@ -2573,17 +2555,15 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[12],zero,zero,zero,zero,ymm4[13],zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,ymm4[18],zero
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm3[26],zero,ymm3[28,u],zero,ymm3[u],zero,ymm3[29],zero,ymm3[31,u],zero,ymm3[30]
+; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm10, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[26],zero,ymm0[28],zero,zero,ymm0[27],zero,ymm0[29],zero,ymm0[31],zero,zero,ymm0[30],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[27,u],zero,ymm1[26],zero,ymm1[28],zero,ymm1[30,u],zero,ymm1[29],zero,ymm1[31,u]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm2, %ymm6, %ymm0
+; AVX512-FCP-NEXT: vpternlogq $226, %ymm2, %ymm6, %ymm0
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [6,6,6,0,7,7,7,7]
; AVX512-FCP-NEXT: vpermd %ymm4, %ymm1, %ymm1
; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
@@ -2595,10 +2575,10 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQ-LABEL: store_i8_stride5_vf32:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm1
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm2
; AVX512DQ-NEXT: vmovdqa (%r8), %ymm0
; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm5
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm5[8],zero,xmm5[u,7],zero,xmm5[9],zero,xmm5[u],zero,xmm5[u,10],zero,xmm5[12],zero,xmm5[u,11]
@@ -2628,45 +2608,40 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
; AVX512DQ-NEXT: vpermd %zmm6, %zmm8, %zmm6
; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u]
; AVX512DQ-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u],zero,ymm2[13,u,u,u],zero,ymm2[14,u,u,u],zero,ymm2[15,u,u,u],zero,ymm2[16,u,u,u],zero,ymm2[17,u,u,u],zero,ymm2[18,u,u,u],zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,13],zero,ymm1[u,u,u,14],zero,ymm1[u,u,u,15],zero,ymm1[u,u,u,16],zero,ymm1[u,u,u,17],zero,ymm1[u,u,u,18],zero,ymm1[u,u,u,19]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u,u],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u,u,19]
; AVX512DQ-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX512DQ-NEXT: vpternlogq $226, %ymm5, %ymm11, %ymm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21,u],zero,ymm4[20],zero,ymm4[22],zero,ymm4[24,u],zero,ymm4[23],zero,ymm4[25,u]
+; AVX512DQ-NEXT: vpor %ymm5, %ymm9, %ymm5
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512DQ-NEXT: vpor %ymm9, %ymm10, %ymm9
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm10, %ymm5, %ymm9
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm3[21],zero,ymm3[21,20],zero,ymm3[22],zero,ymm3[24],zero,ymm3[22,23],zero,ymm3[25]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512DQ-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm5
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm9
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm5
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[12],zero,zero,zero,zero,ymm0[13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero,ymm0[18],zero
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm0[0,2,1,1,4,6,5,5]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,3,2]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX512DQ-NEXT: vpandn %ymm9, %ymm11, %ymm9
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX512DQ-NEXT: vpandn %ymm9, %ymm10, %ymm9
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,ymm3[26],zero,ymm3[28],zero,ymm3[30],zero,zero,ymm3[29],zero,ymm3[31],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,ymm3[27],zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm4[26],zero,ymm4[28,u],zero,ymm4[u],zero,ymm4[29],zero,ymm4[31,u],zero,ymm4[30]
+; AVX512DQ-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm10, %ymm4, %ymm3
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,ymm1[26],zero,ymm1[28],zero,ymm1[30],zero,zero,ymm1[29],zero,ymm1[31],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27,u],zero,ymm2[26],zero,ymm2[28],zero,ymm2[30,u],zero,ymm2[29],zero,ymm2[31,u]
+; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512DQ-NEXT: vpternlogq $184, %ymm3, %ymm7, %ymm1
+; AVX512DQ-NEXT: vpternlogq $226, %ymm3, %ymm7, %ymm1
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
@@ -2678,10 +2653,10 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQ-FCP-LABEL: store_i8_stride5_vf32:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm4
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm6
@@ -2710,26 +2685,23 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
; AVX512DQ-FCP-NEXT: vpermd %zmm4, %zmm7, %zmm7
; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u]
; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u],zero,ymm1[13,u,u,u],zero,ymm1[14,u,u,u],zero,ymm1[15,u,u,u],zero,ymm1[16,u,u,u],zero,ymm1[17,u,u,u],zero,ymm1[18,u,u,u],zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,13],zero,ymm0[u,u,u,14],zero,ymm0[u,u,u,15],zero,ymm0[u,u,u,16],zero,ymm0[u,u,u,17],zero,ymm0[u,u,u,18],zero,ymm0[u,u,u,19]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u],zero,ymm3[13,u,u,u],zero,ymm3[14,u,u,u],zero,ymm3[15,u,u,u],zero,ymm3[16,u,u,u],zero,ymm3[17,u,u,u],zero,ymm3[18,u,u,u],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,13],zero,ymm2[u,u,u,14],zero,ymm2[u,u,u,15],zero,ymm2[u,u,u,16],zero,ymm2[u,u,u,17],zero,ymm2[u,u,u,18],zero,ymm2[u,u,u,19]
; AVX512DQ-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX512DQ-FCP-NEXT: vpternlogq $226, %ymm5, %ymm10, %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21,u],zero,ymm3[20],zero,ymm3[22],zero,ymm3[24,u],zero,ymm3[23],zero,ymm3[25,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm9, %ymm5
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[21],zero,zero,ymm0[20],zero,ymm0[22],zero,ymm0[24],zero,zero,ymm0[23],zero,ymm0[25],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[19],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512DQ-FCP-NEXT: vpor %ymm9, %ymm10, %ymm9
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm10, %ymm5, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm2[21],zero,ymm2[21,20],zero,ymm2[22],zero,ymm2[24],zero,ymm2[22,23],zero,ymm2[25]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm9
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm5
; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm8[0,1,2,3],zmm5[4,5,6,7]
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [4,0,5,5,5,5,0,6]
; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm8, %ymm8
@@ -2738,17 +2710,15 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[12],zero,zero,zero,zero,ymm4[13],zero,zero,zero,zero,ymm4[14],zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,ymm4[18],zero
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,27,u,u,26,u,28,u,30,u,u,29,u,31,u]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u],zero,ymm3[26],zero,ymm3[28,u],zero,ymm3[u],zero,ymm3[29],zero,ymm3[31,u],zero,ymm3[30]
+; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm10, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[26],zero,ymm0[28],zero,zero,ymm0[27],zero,ymm0[29],zero,ymm0[31],zero,zero,ymm0[30],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[27,u],zero,ymm1[26],zero,ymm1[28],zero,ymm1[30,u],zero,ymm1[29],zero,ymm1[31,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm2, %ymm6, %ymm0
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %ymm2, %ymm6, %ymm0
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [6,6,6,0,7,7,7,7]
; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
@@ -2792,26 +2762,24 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-NEXT: kmovq %rax, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
; AVX512BW-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
; AVX512BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm5[21],zero,zero,ymm5[20],zero,ymm5[22],zero,ymm5[24],zero,zero,ymm5[23],zero,ymm5[25],zero
; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,3,3,0,4,4,4,4]
-; AVX512BW-NEXT: vpermd %ymm4, %ymm8, %ymm8
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
+; AVX512BW-NEXT: vpermd %ymm4, %ymm7, %ymm7
; AVX512BW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512BW-NEXT: kmovd %eax, %k1
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 {%k1} = ymm5[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm7 {%k1} = ymm5[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm5[21],zero,zero,ymm5[20],zero,ymm5[22],zero,ymm5[24],zero,zero,ymm5[23],zero,ymm5[25],zero
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
+; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512BW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512BW-NEXT: kmovq %rax, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
@@ -2854,11 +2822,11 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-LABEL: store_i8_stride5_vf32:
; AVX512BW-FCP: # %bb.0:
; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm3
; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm0
; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm6
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
; AVX512BW-FCP-NEXT: vpor %xmm5, %xmm7, %xmm5
@@ -2871,39 +2839,37 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[6],zero,xmm7[8],zero,xmm7[u,7],zero,xmm7[9],zero,xmm7[11],zero,xmm7[u,10],zero,xmm7[12],zero
; AVX512BW-FCP-NEXT: vpor %xmm7, %xmm8, %xmm7
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm4, %zmm4
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,0,1,1,4,4,5,5]
; AVX512BW-FCP-NEXT: movabsq $3570337559743967628, %rax # imm = 0x318C631818C6318C
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm4 {%k1}
; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
; AVX512BW-FCP-NEXT: vpermd %zmm5, %zmm6, %zmm6
; AVX512BW-FCP-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1}
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
; AVX512BW-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
; AVX512BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,3,3,0,4,4,4,4]
-; AVX512BW-FCP-NEXT: vpermd %ymm1, %ymm8, %ymm8
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
+; AVX512BW-FCP-NEXT: vpermd %ymm1, %ymm7, %ymm7
; AVX512BW-FCP-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 {%k1} = ymm4[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 {%k1} = ymm3[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512BW-FCP-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
@@ -2912,16 +2878,14 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,zero,zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm1, %ymm1
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm1, %ymm1
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512BW-FCP-NEXT: movl $415641996, %eax # imm = 0x18C6318C
; AVX512BW-FCP-NEXT: kmovd %eax, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm1, %ymm0 {%k1}
@@ -2932,7 +2896,7 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm1, %ymm0 {%k1}
; AVX512BW-FCP-NEXT: vmovdqa %ymm0, 128(%r9)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%r9)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, (%r9)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
@@ -2970,26 +2934,24 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
; AVX512DQ-BW-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm1[12,13],zero,zero,zero,zero,ymm1[14],zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,ymm1[16],zero,zero,zero,ymm1[16,17],zero,zero,zero,zero,ymm1[18],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm1[21],zero,ymm1[21,20],zero,ymm1[22],zero,ymm1[24],zero,ymm1[22,23],zero,ymm1[25]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm5[21],zero,zero,ymm5[20],zero,ymm5[22],zero,ymm5[24],zero,zero,ymm5[23],zero,ymm5[25],zero
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,3,3,0,4,4,4,4]
-; AVX512DQ-BW-NEXT: vpermd %ymm4, %ymm8, %ymm8
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
+; AVX512DQ-BW-NEXT: vpermd %ymm4, %ymm7, %ymm7
; AVX512DQ-BW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512DQ-BW-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 {%k1} = ymm5[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm7 {%k1} = ymm5[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm5[21],zero,zero,ymm5[20],zero,ymm5[22],zero,ymm5[24],zero,zero,ymm5[23],zero,ymm5[25],zero
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512DQ-BW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
@@ -3032,11 +2994,11 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-LABEL: store_i8_stride5_vf32:
; AVX512DQ-BW-FCP: # %bb.0:
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm3
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm0
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[u],zero,xmm4[u,10],zero,xmm4[12],zero,xmm4[u,11]
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm6
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9,u,11,u],zero,xmm6[10],zero,xmm6[12,u],zero
; AVX512DQ-BW-FCP-NEXT: vpor %xmm5, %xmm7, %xmm5
@@ -3049,39 +3011,37 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm8[6],zero,xmm8[8,u],zero,xmm8[7],zero,xmm8[9],zero,xmm8[11,u],zero,xmm8[10],zero,xmm8[12]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[6],zero,xmm7[8],zero,xmm7[u,7],zero,xmm7[9],zero,xmm7[11],zero,xmm7[u,10],zero,xmm7[12],zero
; AVX512DQ-BW-FCP-NEXT: vpor %xmm7, %xmm8, %xmm7
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm4, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,0,1,1,4,4,5,5]
; AVX512DQ-BW-FCP-NEXT: movabsq $3570337559743967628, %rax # imm = 0x318C631818C6318C
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm4 {%k1}
; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
; AVX512DQ-BW-FCP-NEXT: vpermd %zmm5, %zmm6, %zmm6
; AVX512DQ-BW-FCP-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm4 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
; AVX512DQ-BW-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = zero,zero,ymm0[12,13],zero,zero,zero,zero,ymm0[14],zero,zero,zero,ymm0[14,15],zero,zero,zero,zero,ymm0[16],zero,zero,zero,ymm0[16,17],zero,zero,zero,zero,ymm0[18],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,ymm2[13],zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,ymm2[18],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19],zero,ymm0[21],zero,ymm0[21,20],zero,ymm0[22],zero,ymm0[24],zero,ymm0[22,23],zero,ymm0[25]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[19],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero
; AVX512DQ-BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [3,3,3,0,4,4,4,4]
-; AVX512DQ-BW-FCP-NEXT: vpermd %ymm1, %ymm8, %ymm8
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [3,3,3,0,4,4,4,4]
+; AVX512DQ-BW-FCP-NEXT: vpermd %ymm1, %ymm7, %ymm7
; AVX512DQ-BW-FCP-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 {%k1} = ymm4[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm7 {%k1} = ymm3[u,u,13,u,u,u,u,14,u,u,u,u,15,u,u,u,u,16,u,u,u,u,17,u,u,u,u,18,u,u,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19],zero,ymm3[21],zero,zero,ymm3[20],zero,ymm3[22],zero,ymm3[24],zero,zero,ymm3[23],zero,ymm3[25],zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512DQ-BW-FCP-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
@@ -3090,16 +3050,14 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm7 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,zero,zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm3[26],zero,ymm3[28],zero,zero,zero,zero,ymm3[29],zero,ymm3[31],zero,zero,ymm3[30]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm1, %ymm1
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[27],zero,zero,ymm2[26],zero,ymm2[28],zero,ymm2[30],zero,zero,ymm2[29],zero,ymm2[31],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,zero,ymm0[26],zero,ymm0[28],zero,ymm0[30],zero,zero,ymm0[29],zero,ymm0[31],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512DQ-BW-FCP-NEXT: movl $415641996, %eax # imm = 0x18C6318C
; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm1, %ymm0 {%k1}
@@ -3110,7 +3068,7 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm1, %ymm0 {%k1}
; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, 128(%r9)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 64(%r9)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, (%r9)
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
@@ -4148,209 +4106,200 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX2-LABEL: store_i8_stride5_vf64:
; AVX2: # %bb.0:
-; AVX2-NEXT: subq $312, %rsp # imm = 0x138
-; AVX2-NEXT: vmovdqa 32(%rdi), %ymm10
+; AVX2-NEXT: subq $248, %rsp
+; AVX2-NEXT: vmovdqa 32(%rdi), %ymm4
; AVX2-NEXT: vmovdqa (%rcx), %xmm1
; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-NEXT: vmovdqa 32(%rcx), %xmm11
-; AVX2-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX2-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT: vmovdqa {{.*#+}} xmm0 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
; AVX2-NEXT: vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT: vmovdqa (%rdx), %xmm3
; AVX2-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-NEXT: vmovdqa 32(%rdx), %xmm12
-; AVX2-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX2-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT: vmovdqa {{.*#+}} xmm2 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
; AVX2-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX2-NEXT: vpor %xmm1, %xmm3, %xmm1
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX2-NEXT: vmovdqa (%rdi), %xmm4
-; AVX2-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT: vmovdqa (%rdi), %xmm5
+; AVX2-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX2-NEXT: vpshufb %xmm3, %xmm4, %xmm4
-; AVX2-NEXT: vmovdqa (%rsi), %xmm14
-; AVX2-NEXT: vmovdqa {{.*#+}} xmm5 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX2-NEXT: vpshufb %xmm5, %xmm14, %xmm8
-; AVX2-NEXT: vpor %xmm4, %xmm8, %xmm4
-; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm8 = [255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255]
-; AVX2-NEXT: vpblendvb %ymm8, %ymm1, %ymm4, %ymm1
-; AVX2-NEXT: vmovdqa (%r8), %xmm4
-; AVX2-NEXT: vmovdqa %xmm4, (%rsp) # 16-byte Spill
-; AVX2-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,2]
-; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,1,1]
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
-; AVX2-NEXT: vpblendvb %ymm9, %ymm1, %ymm4, %ymm1
+; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm5
+; AVX2-NEXT: vmovdqa (%rsi), %xmm6
+; AVX2-NEXT: vmovdqa {{.*#+}} xmm8 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX2-NEXT: vpshufb %xmm8, %xmm6, %xmm9
+; AVX2-NEXT: vpor %xmm5, %xmm9, %xmm5
+; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255]
+; AVX2-NEXT: vpblendvb %ymm9, %ymm1, %ymm5, %ymm1
+; AVX2-NEXT: vmovdqa (%r8), %xmm5
+; AVX2-NEXT: vmovdqa %xmm5, (%rsp) # 16-byte Spill
+; AVX2-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
+; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,1,1]
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
+; AVX2-NEXT: vpblendvb %ymm12, %ymm1, %ymm5, %ymm1
; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX2-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-NEXT: vpshufb %xmm0, %xmm11, %xmm0
-; AVX2-NEXT: vpshufb %xmm2, %xmm12, %xmm1
+; AVX2-NEXT: vmovdqa 32(%rdi), %xmm5
+; AVX2-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT: vpshufb %xmm0, %xmm7, %xmm0
+; AVX2-NEXT: vpshufb %xmm2, %xmm10, %xmm1
; AVX2-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX2-NEXT: vmovdqa 32(%rsi), %xmm2
; AVX2-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-NEXT: vpshufb %xmm3, %xmm4, %xmm1
-; AVX2-NEXT: vpshufb %xmm5, %xmm2, %xmm2
+; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm1
+; AVX2-NEXT: vpshufb %xmm8, %xmm2, %xmm2
; AVX2-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX2-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX2-NEXT: vmovdqa 32(%rsi), %ymm11
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX2-NEXT: vpblendvb %ymm8, %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpblendvb %ymm9, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vmovdqa 32(%r8), %xmm1
; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,2]
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,1,1]
-; AVX2-NEXT: vpblendvb %ymm9, %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpblendvb %ymm12, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,20,128,22,128,24,128,22,23,128,25,128,23]
-; AVX2-NEXT: vpshufb %ymm15, %ymm10, %ymm1
-; AVX2-NEXT: vmovdqa %ymm10, %ymm11
-; AVX2-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
-; AVX2-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX2-NEXT: vpshufb %ymm5, %ymm2, %ymm3
-; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX2-NEXT: vpshufb %ymm15, %ymm4, %ymm1
+; AVX2-NEXT: vmovdqa %ymm4, %ymm13
+; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
+; AVX2-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX2-NEXT: vpshufb %ymm4, %ymm11, %ymm3
; AVX2-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX2-NEXT: vmovdqa 32(%rcx), %ymm7
+; AVX2-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX2-NEXT: vmovdqa 32(%rcx), %ymm14
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX2-NEXT: vpshufb %ymm3, %ymm7, %ymm4
-; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm4[2,2,3,3]
-; AVX2-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25]
-; AVX2-NEXT: vpshufb %ymm4, %ymm13, %ymm12
-; AVX2-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX2-NEXT: vpor %ymm9, %ymm12, %ymm9
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm12 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
-; AVX2-NEXT: vpblendvb %ymm12, %ymm1, %ymm9, %ymm9
-; AVX2-NEXT: vmovdqa (%rdi), %ymm6
-; AVX2-NEXT: vpshufb %ymm15, %ymm6, %ymm1
-; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpshufb %ymm3, %ymm14, %ymm8
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25]
+; AVX2-NEXT: vpshufb %ymm5, %ymm12, %ymm10
+; AVX2-NEXT: vpor %ymm8, %ymm10, %ymm8
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm10 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
+; AVX2-NEXT: vpblendvb %ymm10, %ymm1, %ymm8, %ymm2
+; AVX2-NEXT: vmovdqa (%rdi), %ymm9
+; AVX2-NEXT: vpshufb %ymm15, %ymm9, %ymm1
; AVX2-NEXT: vmovdqa (%rsi), %ymm15
-; AVX2-NEXT: vpshufb %ymm5, %ymm15, %ymm5
-; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX2-NEXT: vpor %ymm1, %ymm5, %ymm5
-; AVX2-NEXT: vmovdqa (%rcx), %ymm10
-; AVX2-NEXT: vpshufb %ymm3, %ymm10, %ymm3
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm3[2,2,3,3]
-; AVX2-NEXT: vmovdqa (%rdx), %ymm8
-; AVX2-NEXT: vpshufb %ymm4, %ymm8, %ymm4
+; AVX2-NEXT: vpshufb %ymm4, %ymm15, %ymm4
+; AVX2-NEXT: vpor %ymm1, %ymm4, %ymm4
+; AVX2-NEXT: vmovdqa (%rcx), %ymm7
+; AVX2-NEXT: vpshufb %ymm3, %ymm7, %ymm0
+; AVX2-NEXT: vmovdqa (%rdx), %ymm3
+; AVX2-NEXT: vpshufb %ymm5, %ymm3, %ymm5
+; AVX2-NEXT: vpor %ymm0, %ymm5, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX2-NEXT: vpor %ymm0, %ymm4, %ymm0
-; AVX2-NEXT: vpblendvb %ymm12, %ymm5, %ymm0, %ymm0
-; AVX2-NEXT: vmovdqa 32(%r8), %ymm12
-; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[0,2,1,1,4,6,5,5]
+; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX2-NEXT: vpblendvb %ymm10, %ymm4, %ymm0, %ymm0
+; AVX2-NEXT: vmovdqa 32(%r8), %ymm10
+; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm10[0,2,1,1,4,6,5,5]
; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,3,2]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX2-NEXT: vpblendvb %ymm5, %ymm9, %ymm4, %ymm1
+; AVX2-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm1
; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vmovdqa (%r8), %ymm9
-; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm9[0,2,1,1,4,6,5,5]
+; AVX2-NEXT: vmovdqa (%r8), %ymm8
+; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[0,2,1,1,4,6,5,5]
; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,3,2]
; AVX2-NEXT: vpblendvb %ymm5, %ymm0, %ymm4, %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpmovsxbd {{.*#+}} ymm3 = [3,3,3,0,4,4,4,4]
-; AVX2-NEXT: vpermd %ymm11, %ymm3, %ymm4
+; AVX2-NEXT: vpmovsxbd {{.*#+}} ymm2 = [3,3,3,0,4,4,4,4]
+; AVX2-NEXT: vpermd %ymm13, %ymm2, %ymm4
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm5 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
-; AVX2-NEXT: vpshufb %ymm5, %ymm2, %ymm0
+; AVX2-NEXT: vpshufb %ymm5, %ymm11, %ymm0
; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255]
; AVX2-NEXT: vpblendvb %ymm1, %ymm4, %ymm0, %ymm0
-; AVX2-NEXT: vpermd %ymm6, %ymm3, %ymm2
+; AVX2-NEXT: vpermd %ymm9, %ymm2, %ymm2
; AVX2-NEXT: vpshufb %ymm5, %ymm15, %ymm4
; AVX2-NEXT: vpblendvb %ymm1, %ymm2, %ymm4, %ymm1
; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX2-NEXT: vmovdqa %ymm7, %ymm3
-; AVX2-NEXT: vpshufb %ymm2, %ymm7, %ymm4
+; AVX2-NEXT: vpshufb %ymm2, %ymm14, %ymm4
; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128]
-; AVX2-NEXT: vpshufb %ymm5, %ymm13, %ymm11
-; AVX2-NEXT: vpor %ymm4, %ymm11, %ymm4
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm11 = [u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255]
-; AVX2-NEXT: vpblendvb %ymm11, %ymm0, %ymm4, %ymm0
-; AVX2-NEXT: vpshufb %ymm2, %ymm10, %ymm2
-; AVX2-NEXT: vpshufb %ymm5, %ymm8, %ymm4
+; AVX2-NEXT: vpshufb %ymm5, %ymm12, %ymm13
+; AVX2-NEXT: vpor %ymm4, %ymm13, %ymm4
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm13 = [u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255]
+; AVX2-NEXT: vpblendvb %ymm13, %ymm0, %ymm4, %ymm0
+; AVX2-NEXT: vpshufb %ymm2, %ymm7, %ymm2
+; AVX2-NEXT: vpshufb %ymm5, %ymm3, %ymm4
; AVX2-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX2-NEXT: vpblendvb %ymm11, %ymm1, %ymm2, %ymm1
+; AVX2-NEXT: vpblendvb %ymm13, %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vpmovsxbd {{.*#+}} ymm2 = [3,3,3,3,0,4,4,4]
-; AVX2-NEXT: vpermd %ymm12, %ymm2, %ymm4
+; AVX2-NEXT: vpermd %ymm10, %ymm2, %ymm4
; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
; AVX2-NEXT: vpblendvb %ymm5, %ymm0, %ymm4, %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpermd %ymm9, %ymm2, %ymm0
+; AVX2-NEXT: vpermd %ymm8, %ymm2, %ymm0
; AVX2-NEXT: vpblendvb %ymm5, %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3],xmm0[4],xmm14[4],xmm0[5],xmm14[5],xmm0[6],xmm14[6],xmm0[7],xmm14[7]
+; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX2-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX2-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX2-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX2-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255]
-; AVX2-NEXT: vpblendvb %ymm11, %ymm0, %ymm1, %ymm6
-; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX2-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX2-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX2-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX2-NEXT: vpshufb %xmm13, %xmm0, %xmm0
+; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,1,1]
+; AVX2-NEXT: vmovdqa {{.*#+}} xmm6 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX2-NEXT: vpshufb %xmm6, %xmm1, %xmm0
+; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm0[0,0,1,1]
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm0 = [255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255]
+; AVX2-NEXT: vpblendvb %ymm0, %ymm2, %ymm1, %ymm4
; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX2-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
+; AVX2-NEXT: vpshufb %xmm13, %xmm1, %xmm1
+; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
+; AVX2-NEXT: # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
+; AVX2-NEXT: vpshufb %xmm6, %xmm2, %xmm2
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX2-NEXT: vpblendvb %ymm11, %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,1]
+; AVX2-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
; AVX2-NEXT: vpshufd $80, (%rsp), %xmm1 # 16-byte Folded Reload
; AVX2-NEXT: # xmm1 = mem[0,0,1,1]
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255]
-; AVX2-NEXT: vpblendvb %ymm2, %ymm6, %ymm1, %ymm6
+; AVX2-NEXT: vpblendvb %ymm2, %ymm4, %ymm1, %ymm4
; AVX2-NEXT: vpshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; AVX2-NEXT: # xmm1 = mem[0,0,1,1]
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
-; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm7
+; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm6
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm1
-; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm13[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
+; AVX2-NEXT: vpshufb %ymm0, %ymm14, %ymm1
+; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm12[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,3,3,6,6,7,7]
-; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [255,0,255,0,0,255,0,255,0,255,0,0,255,0,255,0,255,0,255,0,0,255,0,255,0,255,0,0,255,0,255,0]
-; AVX2-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX2-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1
-; AVX2-NEXT: vpshufb %ymm0, %ymm10, %ymm0
-; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm8[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
+; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [255,0,255,0,0,255,0,255,0,255,0,0,255,0,255,0,255,0,255,0,0,255,0,255,0,255,0,0,255,0,255,0]
+; AVX2-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX2-NEXT: vpblendvb %ymm5, %ymm1, %ymm2, %ymm1
+; AVX2-NEXT: vpshufb %ymm0, %ymm7, %ymm0
+; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm3[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,3,3,6,6,7,7]
-; AVX2-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0
+; AVX2-NEXT: vpblendvb %ymm5, %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
-; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-NEXT: vpshufb %ymm2, %ymm3, %ymm3
-; AVX2-NEXT: vpshufhw $230, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX2-NEXT: # ymm4 = mem[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
-; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
-; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [255,0,0,255,0,255,0,0,0,0,255,0,255,0,0,255,255,0,0,255,0,255,0,0,0,0,255,0,255,0,0,255]
-; AVX2-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX2-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm3
+; AVX2-NEXT: vpshufb %ymm2, %ymm11, %ymm3
+; AVX2-NEXT: vpshufhw $230, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX2-NEXT: # ymm5 = mem[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
+; AVX2-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,3,3,6,6,7,7]
+; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [255,0,0,255,0,255,0,0,0,0,255,0,255,0,0,255,255,0,0,255,0,255,0,0,0,0,255,0,255,0,0,255]
+; AVX2-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX2-NEXT: vpblendvb %ymm7, %ymm3, %ymm5, %ymm3
; AVX2-NEXT: vpshufb %ymm2, %ymm15, %ymm2
-; AVX2-NEXT: vpshufhw $230, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX2-NEXT: # ymm4 = mem[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
-; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
-; AVX2-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2
+; AVX2-NEXT: vpshufhw {{.*#+}} ymm5 = ymm9[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
+; AVX2-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,3,3,6,6,7,7]
+; AVX2-NEXT: vpblendvb %ymm7, %ymm2, %ymm5, %ymm2
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm4 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
-; AVX2-NEXT: vpblendvb %ymm4, %ymm1, %ymm3, %ymm1
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm5 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
+; AVX2-NEXT: vpblendvb %ymm5, %ymm1, %ymm3, %ymm1
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX2-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0
-; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[2,2,3,3,6,6,7,7]
+; AVX2-NEXT: vpblendvb %ymm5, %ymm0, %ymm2, %ymm0
+; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm10[2,2,3,3,6,6,7,7]
; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
; AVX2-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1
-; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm9[2,2,3,3,6,6,7,7]
+; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[2,2,3,3,6,6,7,7]
; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX2-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
@@ -4362,21 +4311,21 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-NEXT: vmovaps %ymm2, 256(%r9)
; AVX2-NEXT: vmovdqa %ymm0, 128(%r9)
-; AVX2-NEXT: vmovdqa %ymm7, 160(%r9)
+; AVX2-NEXT: vmovdqa %ymm6, 160(%r9)
; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT: vmovaps %ymm0, 192(%r9)
; AVX2-NEXT: vmovdqa %ymm1, 288(%r9)
-; AVX2-NEXT: vmovdqa %ymm6, (%r9)
+; AVX2-NEXT: vmovdqa %ymm4, (%r9)
; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT: vmovaps %ymm0, 32(%r9)
-; AVX2-NEXT: addq $312, %rsp # imm = 0x138
+; AVX2-NEXT: addq $248, %rsp
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
;
; AVX2-FP-LABEL: store_i8_stride5_vf64:
; AVX2-FP: # %bb.0:
; AVX2-FP-NEXT: subq $200, %rsp
-; AVX2-FP-NEXT: vmovdqa 32(%rdx), %ymm11
+; AVX2-FP-NEXT: vmovdqa 32(%rdx), %ymm12
; AVX2-FP-NEXT: vmovdqa (%rcx), %xmm1
; AVX2-FP-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX2-FP-NEXT: vmovdqa 32(%rcx), %xmm8
@@ -4420,7 +4369,7 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vpshufb %xmm3, %xmm4, %xmm1
; AVX2-FP-NEXT: vpshufb %xmm5, %xmm2, %xmm2
; AVX2-FP-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX2-FP-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX2-FP-NEXT: vmovdqa 32(%rcx), %ymm2
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm1, %ymm0
@@ -4431,106 +4380,98 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vpblendvb %ymm7, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,29,26,128,28,128,30,128,28,29,128,31,128,29]
-; AVX2-FP-NEXT: vpshufb %ymm0, %ymm11, %ymm1
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX2-FP-NEXT: vpshufb %ymm0, %ymm12, %ymm1
; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
; AVX2-FP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm3, %ymm2
-; AVX2-FP-NEXT: vmovdqa %ymm3, %ymm12
-; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX2-FP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-FP-NEXT: vpshufb %ymm8, %ymm2, %ymm3
+; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm14
+; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX2-FP-NEXT: vmovdqa 32(%rsi), %ymm11
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,29,26,128,28,128,26,27,28,29,128,31,128,29,30,128]
-; AVX2-FP-NEXT: vpshufb %ymm5, %ymm4, %ymm2
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm2[2,2,3,3]
-; AVX2-FP-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX2-FP-NEXT: vpshufb %ymm5, %ymm4, %ymm6
; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
; AVX2-FP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm2, %ymm7
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX2-FP-NEXT: vpshufb %ymm9, %ymm11, %ymm7
; AVX2-FP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm1, %ymm6, %ymm3
; AVX2-FP-NEXT: vmovdqa (%rdx), %ymm13
; AVX2-FP-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa (%rcx), %ymm7
; AVX2-FP-NEXT: vpshufb %ymm8, %ymm7, %ymm1
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm8
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX2-FP-NEXT: vpshufb %ymm5, %ymm1, %ymm0
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm0[2,2,3,3]
-; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm0
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm0, %ymm9
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX2-FP-NEXT: vpshufb %ymm5, %ymm2, %ymm5
+; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX2-FP-NEXT: vpshufb %ymm9, %ymm1, %ymm9
; AVX2-FP-NEXT: vpor %ymm5, %ymm9, %ymm5
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm8, %ymm5, %ymm8
; AVX2-FP-NEXT: vmovdqa 32(%r8), %ymm5
; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,3,3,6,6,7,7]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm3, %ymm9, %ymm3
-; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm3, %ymm9, %ymm0
+; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa (%r8), %ymm3
; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm9 = ymm3[2,2,3,3,6,6,7,7]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm6
-; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm0
+; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,20,128,22,128,24,128,22,23,128,25,128,23]
; AVX2-FP-NEXT: vpshufb %ymm8, %ymm4, %ymm9
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
; AVX2-FP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm2, %ymm15
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm11, %ymm15
; AVX2-FP-NEXT: vpor %ymm9, %ymm15, %ymm9
; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX2-FP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vpshufb %ymm15, %ymm12, %ymm14
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
+; AVX2-FP-NEXT: vpshufb %ymm15, %ymm14, %ymm0
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25]
-; AVX2-FP-NEXT: vpshufb %ymm6, %ymm11, %ymm12
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX2-FP-NEXT: vpor %ymm14, %ymm12, %ymm12
+; AVX2-FP-NEXT: vpshufb %ymm6, %ymm12, %ymm14
+; AVX2-FP-NEXT: vpor %ymm0, %ymm14, %ymm0
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm14 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
-; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm9, %ymm12, %ymm9
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm1, %ymm8
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm0, %ymm10
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
-; AVX2-FP-NEXT: vpor %ymm8, %ymm10, %ymm8
-; AVX2-FP-NEXT: vpshufb %ymm15, %ymm7, %ymm10
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
+; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm9, %ymm0, %ymm0
+; AVX2-FP-NEXT: vpshufb %ymm8, %ymm2, %ymm8
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm1, %ymm9
+; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FP-NEXT: vpshufb %ymm15, %ymm7, %ymm9
; AVX2-FP-NEXT: vpshufb %ymm6, %ymm13, %ymm6
+; AVX2-FP-NEXT: vpor %ymm6, %ymm9, %ymm6
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX2-FP-NEXT: vpor %ymm6, %ymm10, %ymm6
; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm8, %ymm6, %ymm6
; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm8 = ymm5[0,2,1,1,4,6,5,5]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,3,2]
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm9, %ymm8, %ymm10
-; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm8 = ymm3[0,2,1,1,4,6,5,5]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,3,2]
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm6, %ymm8, %ymm9
-; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [3,3,3,0,4,4,4,4]
-; AVX2-FP-NEXT: vpermd %ymm4, %ymm6, %ymm4
-; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} ymm8 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm2, %ymm2
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255]
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm4, %ymm2, %ymm2
-; AVX2-FP-NEXT: vpermd %ymm1, %ymm6, %ymm1
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm0, %ymm0
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm1, %ymm0, %ymm0
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm0, %ymm8, %ymm10
+; AVX2-FP-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[0,2,1,1,4,6,5,5]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
+; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm6, %ymm0, %ymm9
+; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [3,3,3,0,4,4,4,4]
+; AVX2-FP-NEXT: vpermd %ymm4, %ymm0, %ymm4
+; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX2-FP-NEXT: vpshufb %ymm6, %ymm11, %ymm8
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255]
+; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm4, %ymm8, %ymm4
+; AVX2-FP-NEXT: vpermd %ymm2, %ymm0, %ymm0
+; AVX2-FP-NEXT: vpshufb %ymm6, %ymm1, %ymm1
+; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm1, %ymm4, %ymm4
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm1, %ymm2, %ymm2
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128]
-; AVX2-FP-NEXT: vpshufb %ymm6, %ymm11, %ymm8
-; AVX2-FP-NEXT: vpor %ymm4, %ymm8, %ymm4
+; AVX2-FP-NEXT: vpshufb %ymm6, %ymm12, %ymm8
+; AVX2-FP-NEXT: vpor %ymm2, %ymm8, %ymm2
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255]
-; AVX2-FP-NEXT: vpblendvb %ymm8, %ymm2, %ymm4, %ymm2
+; AVX2-FP-NEXT: vpblendvb %ymm8, %ymm4, %ymm2, %ymm2
; AVX2-FP-NEXT: vpshufb %ymm1, %ymm7, %ymm1
; AVX2-FP-NEXT: vpshufb %ymm6, %ymm13, %ymm4
; AVX2-FP-NEXT: vpor %ymm1, %ymm4, %ymm1
@@ -4596,8 +4537,8 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-LABEL: store_i8_stride5_vf64:
; AVX2-FCP: # %bb.0:
; AVX2-FCP-NEXT: subq $168, %rsp
-; AVX2-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX2-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
+; AVX2-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX2-FCP-NEXT: vmovdqa 32(%rcx), %ymm9
; AVX2-FCP-NEXT: vmovdqa (%r8), %ymm11
; AVX2-FCP-NEXT: vmovdqa (%rcx), %xmm1
; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -4643,96 +4584,88 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm1, %ymm2, %ymm1
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vmovdqa 32(%r8), %ymm12
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm3, %ymm1
+; AVX2-FCP-NEXT: vmovdqa 32(%r8), %ymm13
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm3, %ymm1
; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,29,26,128,28,128,30,128,28,29,128,31,128,29]
-; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm1
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
-; AVX2-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm10, %ymm2
-; AVX2-FCP-NEXT: vmovdqa %ymm10, %ymm14
-; AVX2-FCP-NEXT: vmovdqu %ymm10, (%rsp) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm1
+; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
+; AVX2-FCP-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm2
+; AVX2-FCP-NEXT: vmovdqa %ymm9, %ymm15
+; AVX2-FCP-NEXT: vmovdqu %ymm9, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,29,26,128,28,128,26,27,28,29,128,31,128,29,30,128]
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm2
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[2,2,3,3]
-; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
-; AVX2-FCP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm5
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm4
+; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
+; AVX2-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm5
; AVX2-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u,0,0,255,255,u]
; AVX2-FCP-NEXT: vpblendvb %ymm7, %ymm1, %ymm4, %ymm4
-; AVX2-FCP-NEXT: vmovdqa (%rdx), %ymm10
-; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm0
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX2-FCP-NEXT: vmovdqa (%rdx), %ymm12
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm12, %ymm0
; AVX2-FCP-NEXT: vmovdqa (%rcx), %ymm5
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm1
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
-; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm8
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm1
+; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm10
; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm0
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm0[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm9
; AVX2-FCP-NEXT: vmovdqa (%rsi), %ymm0
-; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm15
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
-; AVX2-FCP-NEXT: vpor %ymm9, %ymm15, %ymm9
-; AVX2-FCP-NEXT: vpblendvb %ymm7, %ymm8, %ymm9, %ymm7
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm8
+; AVX2-FCP-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm10[2,2,3,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX2-FCP-NEXT: vpblendvb %ymm7, %ymm9, %ymm8, %ymm7
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [6,6,6,6,7,7,7,7]
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm8, %ymm9
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm4, %ymm9, %ymm4
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm8, %ymm9
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm4, %ymm9, %ymm4
; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermd %ymm11, %ymm8, %ymm4
-; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm7, %ymm4, %ymm4
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm7, %ymm4, %ymm4
; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,20,128,22,128,24,128,22,23,128,25,128,23]
-; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm7
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm7
; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
; AVX2-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm9
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm9
; AVX2-FCP-NEXT: vpor %ymm7, %ymm9, %ymm7
; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX2-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm14, %ymm15
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm10
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm14
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX2-FCP-NEXT: vpor %ymm15, %ymm14, %ymm14
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm15
+; AVX2-FCP-NEXT: vpor %ymm10, %ymm15, %ymm10
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,3,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0]
-; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm7, %ymm14, %ymm7
+; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm7, %ymm10, %ymm7
; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm6
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm8
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX2-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm8
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm4
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm4
; AVX2-FCP-NEXT: vpor %ymm4, %ymm8, %ymm4
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm6, %ymm4, %ymm4
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [4,6,5,5,5,5,4,6]
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm6, %ymm8
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm14, %ymm7, %ymm8, %ymm9
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm6, %ymm8
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm7, %ymm8, %ymm9
; AVX2-FCP-NEXT: vpermd %ymm11, %ymm6, %ymm6
-; AVX2-FCP-NEXT: vpblendvb %ymm14, %ymm4, %ymm6, %ymm7
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm4, %ymm6, %ymm7
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [3,3,3,0,4,4,4,4]
-; AVX2-FCP-NEXT: vpermd %ymm3, %ymm4, %ymm3
+; AVX2-FCP-NEXT: vpermd %ymm2, %ymm4, %ymm2
; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
-; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255,0,u,u,u,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm3, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm2, %ymm3, %ymm2
; AVX2-FCP-NEXT: vpermd %ymm1, %ymm4, %ymm1
; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm0
; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm1, %ymm0, %ymm0
@@ -4740,16 +4673,16 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqu (%rsp), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128]
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm6
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm6
; AVX2-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255,255,0,0,u,255]
; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm2, %ymm3, %ymm2
; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm3
; AVX2-FCP-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm1, %ymm1
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [3,3,3,3,0,4,4,4]
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm3, %ymm0
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm3, %ymm0
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX2-FCP-NEXT: vpermd %ymm11, %ymm3, %ymm2
@@ -4783,7 +4716,7 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vpermd %ymm11, %ymm4, %ymm5
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm2, %ymm5, %ymm2
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm4, %ymm4
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm4, %ymm4
; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm3, %ymm4, %ymm3
; AVX2-FCP-NEXT: vmovdqa %ymm1, 64(%r9)
; AVX2-FCP-NEXT: vmovdqa %ymm0, 224(%r9)
@@ -4805,766 +4738,740 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-LABEL: store_i8_stride5_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm15, %ymm3, %ymm0
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm20
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
+; AVX512-NEXT: vpshufb %xmm1, %xmm12, %xmm0
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm28
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512-NEXT: vpshufb %xmm2, %xmm10, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm2, %xmm29
+; AVX512-NEXT: vporq %xmm0, %xmm1, %xmm21
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512-NEXT: vpshufb %ymm8, %ymm15, %ymm0
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm13
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512-NEXT: vpshufb %ymm3, %ymm13, %ymm1
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512-NEXT: vpshufb %xmm1, %xmm6, %xmm0
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm30
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512-NEXT: vpshufb %xmm2, %xmm7, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm2, %xmm31
+; AVX512-NEXT: vporq %xmm0, %xmm1, %xmm23
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
+; AVX512-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm9, %ymm5, %ymm0
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
+; AVX512-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm4, %ymm5, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm5, %ymm11, %ymm1
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm2, %ymm11, %ymm11
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm26
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm11, %ymm13, %ymm1
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm0, %ymm15, %ymm14
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm14, %zmm25
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm1, %ymm15, %ymm14
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm15, %ymm13, %ymm13
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm27
+; AVX512-NEXT: vmovdqa (%rcx), %ymm13
+; AVX512-NEXT: vpshufb %ymm8, %ymm13, %ymm8
+; AVX512-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512-NEXT: vporq %ymm8, %ymm3, %ymm16
+; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512-NEXT: vpshufb %ymm15, %ymm14, %ymm3
+; AVX512-NEXT: vporq %ymm0, %ymm3, %ymm17
+; AVX512-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512-NEXT: vpshufb %ymm15, %ymm8, %ymm15
+; AVX512-NEXT: vporq %ymm0, %ymm15, %ymm18
+; AVX512-NEXT: vpshufb %ymm4, %ymm8, %ymm0
+; AVX512-NEXT: vpshufb %ymm2, %ymm3, %ymm2
+; AVX512-NEXT: vporq %ymm0, %ymm2, %ymm19
+; AVX512-NEXT: vpshufb %ymm11, %ymm14, %ymm0
+; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm5 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512-NEXT: vpshufb %xmm5, %xmm1, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm16
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm14 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512-NEXT: vpshufb %xmm14, %xmm4, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm8
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512-NEXT: vpshufb %ymm0, %ymm8, %ymm4
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm11
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512-NEXT: vpshufb %ymm1, %ymm11, %ymm10
-; AVX512-NEXT: vpor %ymm4, %ymm10, %ymm4
-; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm13
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512-NEXT: vpshufb %xmm6, %xmm13, %xmm4
-; AVX512-NEXT: vmovdqa64 %xmm6, %xmm25
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512-NEXT: vpshufb %xmm6, %xmm10, %xmm12
-; AVX512-NEXT: vmovdqa64 %xmm6, %xmm26
-; AVX512-NEXT: vporq %xmm4, %xmm12, %xmm20
-; AVX512-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm22
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm23
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,ymm11[26],zero,ymm11[28],zero,ymm11[30],zero,zero,ymm11[29],zero,ymm11[31],zero,zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm8[19],zero,ymm8[21],zero,zero,ymm8[20],zero,ymm8[22],zero,ymm8[24],zero,zero,ymm8[23],zero
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm24
-; AVX512-NEXT: vmovdqa (%rcx), %ymm12
-; AVX512-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512-NEXT: vmovdqa (%rdx), %ymm6
-; AVX512-NEXT: vpshufb %ymm1, %ymm6, %ymm1
-; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm19
-; AVX512-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512-NEXT: vpshufb %ymm15, %ymm7, %ymm2
-; AVX512-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512-NEXT: vpshufb %ymm9, %ymm4, %ymm3
-; AVX512-NEXT: vporq %ymm2, %ymm3, %ymm21
-; AVX512-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512-NEXT: vpshufb %xmm5, %xmm0, %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-NEXT: vpshufb %xmm14, %xmm5, %xmm9
-; AVX512-NEXT: vporq %xmm3, %xmm9, %xmm27
-; AVX512-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512-NEXT: vpshufb %xmm1, %xmm9, %xmm15
-; AVX512-NEXT: vporq %xmm0, %xmm15, %xmm29
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = mem[1,1,2,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,1]
+; AVX512-NEXT: vpshufb %ymm9, %ymm8, %ymm1
+; AVX512-NEXT: vpshufb %ymm5, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512-NEXT: vmovdqa (%rcx), %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm5, %xmm2
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512-NEXT: vpshufb %xmm3, %xmm9, %xmm3
+; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm4
+; AVX512-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm11
+; AVX512-NEXT: vmovdqa64 %xmm30, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm2
+; AVX512-NEXT: vmovdqa64 %xmm31, %xmm13
+; AVX512-NEXT: vpshufb %xmm13, %xmm3, %xmm13
+; AVX512-NEXT: vpor %xmm2, %xmm13, %xmm13
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512-NEXT: vpshufb %ymm14, %ymm11, %ymm2
+; AVX512-NEXT: vpshufd {{.*#+}} xmm15 = mem[1,1,2,2]
+; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,1,1]
; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm28 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512-NEXT: vpandnq %ymm0, %ymm28, %ymm0
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm15
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512-NEXT: vpshufb %ymm1, %ymm15, %ymm14
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm26
-; AVX512-NEXT: vmovdqa (%r8), %ymm0
-; AVX512-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512-NEXT: vpandnq %ymm0, %ymm30, %ymm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm25
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512-NEXT: vpshufb %ymm0, %ymm8, %ymm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm14 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[19],zero,ymm12[21],zero,zero,ymm12[20],zero,ymm12[22],zero,ymm12[24],zero,zero,ymm12[23],zero
-; AVX512-NEXT: vpshufb %ymm0, %ymm12, %ymm12
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm0, %ymm11, %ymm11
-; AVX512-NEXT: vpshufb %ymm0, %ymm6, %ymm2
-; AVX512-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,ymm6[26],zero,ymm6[28],zero,ymm6[30],zero,zero,ymm6[29],zero,ymm6[31],zero,zero
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3],xmm13[4],xmm10[4],xmm13[5],xmm10[5],xmm13[6],xmm10[6],xmm13[7],xmm10[7]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm13 = ymm7[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,ymm4[27],zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30],zero
-; AVX512-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm4
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [4,0,5,5,5,5,0,6,6,6,6,0,7,7,7,7]
-; AVX512-NEXT: vpermd %zmm15, %zmm1, %zmm31
-; AVX512-NEXT: vmovdqa64 (%r8), %zmm16
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
-; AVX512-NEXT: vpermi2d %zmm15, %zmm16, %zmm1
-; AVX512-NEXT: vmovdqa64 %xmm17, %xmm15
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3],xmm15[4],xmm5[4],xmm15[5],xmm5[5],xmm15[6],xmm5[6],xmm15[7],xmm5[7]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512-NEXT: vpshufb %xmm15, %xmm4, %xmm4
-; AVX512-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512-NEXT: vinserti32x4 $2, %xmm27, %zmm5, %zmm5
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512-NEXT: vmovdqa64 %xmm18, %xmm15
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3],xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-NEXT: vpshufb %xmm15, %xmm10, %xmm10
+; AVX512-NEXT: vpandnq %ymm15, %ymm28, %ymm15
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm15, %zmm2
+; AVX512-NEXT: vmovdqa (%r8), %ymm15
+; AVX512-NEXT: vpshufb %ymm14, %ymm15, %ymm14
+; AVX512-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,2,1,1,4,6,5,5]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,3,2]
+; AVX512-NEXT: vpandnq %ymm15, %ymm29, %ymm15
+; AVX512-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512-NEXT: vpshufb %xmm7, %xmm6, %xmm6
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-NEXT: vpshufb %xmm12, %xmm10, %xmm10
; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512-NEXT: vpshufb %xmm15, %xmm9, %xmm9
-; AVX512-NEXT: vinserti32x4 $2, %xmm29, %zmm9, %zmm9
-; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm15 = mem[0,0,1,1]
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm15 # 32-byte Folded Reload
-; AVX512-NEXT: vpermq {{.*#+}} ymm17 = ymm20[0,0,1,1]
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm18 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512-NEXT: vpternlogq $226, %zmm15, %zmm18, %zmm17
-; AVX512-NEXT: vpternlogq $248, %zmm28, %zmm17, %zmm26
-; AVX512-NEXT: vpermq {{.*#+}} zmm15 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm17 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm20 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400,72056498804555775,72056498804555775,18374967950370078975,18374967950370078975]
-; AVX512-NEXT: vpternlogq $248, %zmm20, %zmm15, %zmm17
-; AVX512-NEXT: vpandq %ymm20, %ymm8, %ymm8
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm8
-; AVX512-NEXT: vpermq {{.*#+}} zmm11 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vporq %zmm11, %zmm8, %zmm8
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-NEXT: vpternlogq $226, %zmm17, %zmm11, %zmm8
-; AVX512-NEXT: vpternlogd $184, %zmm8, %zmm30, %zmm31
-; AVX512-NEXT: vpor %ymm2, %ymm14, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm19, %zmm2
-; AVX512-NEXT: vpternlogq $248, %ymm20, %ymm13, %ymm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm18, %zmm0
-; AVX512-NEXT: vpternlogq $248, %ymm20, %ymm12, %ymm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm2
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm11, %zmm3
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm25
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm1
-; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm9[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512-NEXT: vmovdqa64 (%r8), %zmm15
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm30 = [4,0,5,5,5,5,0,6,6,6,6,0,7,7,7,7]
+; AVX512-NEXT: vpermd %zmm11, %zmm30, %zmm30
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm31 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
+; AVX512-NEXT: vpermi2d %zmm11, %zmm15, %zmm31
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm9[0],xmm5[1],xmm9[1],xmm5[2],xmm9[2],xmm5[3],xmm9[3],xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
+; AVX512-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX512-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
+; AVX512-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX512-NEXT: vinserti32x4 $2, %xmm13, %zmm3, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,0,1,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm20, %zmm5, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm23[0,0,1,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm22, %zmm7, %zmm7
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
+; AVX512-NEXT: vpternlogq $248, %zmm28, %zmm7, %zmm2
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vporq %zmm5, %zmm7, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm9 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm7
+; AVX512-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm30
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm17[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm16, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm19[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm7
+; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm1
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm14
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm31
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm4[0,0,1,1,4,4,5,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512-NEXT: vpermd %zmm16, %zmm0, %zmm0
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512-NEXT: vmovdqa64 %zmm25, 64(%r9)
+; AVX512-NEXT: vpermd %zmm15, %zmm0, %zmm0
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512-NEXT: vmovdqa64 %zmm14, 64(%r9)
; AVX512-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512-NEXT: vmovdqa64 %zmm31, 256(%r9)
-; AVX512-NEXT: vmovdqa64 %zmm26, 192(%r9)
+; AVX512-NEXT: vmovdqa64 %zmm31, 128(%r9)
+; AVX512-NEXT: vmovdqa64 %zmm30, 256(%r9)
+; AVX512-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i8_stride5_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $24, %rsp
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm2
; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm2
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm19
-; AVX512-FCP-NEXT: vpor %xmm0, %xmm2, %xmm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm9, %ymm4
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm12, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm25
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm26
-; AVX512-FCP-NEXT: vpor %xmm2, %xmm4, %xmm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm21
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm22
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm9[27],zero,zero,ymm9[26],zero,ymm9[28],zero,ymm9[30],zero,zero,ymm9[29],zero,ymm9[31],zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[19],zero,ymm7[21],zero,zero,ymm7[20],zero,ymm7[22],zero,ymm7[24],zero,zero,ymm7[23],zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm23
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm14
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm14, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm24
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm20
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm16
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm3, %xmm2
-; AVX512-FCP-NEXT: vporq %xmm1, %xmm2, %xmm28
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512-FCP-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm11
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm17
-; AVX512-FCP-NEXT: vporq %xmm8, %xmm11, %xmm29
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [1,1,2,2,2,2,2,2]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm3
+; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm15[8,u],zero,xmm15[7],zero,xmm15[9,u,11,u],zero,xmm15[10],zero,xmm15[12,u],zero
+; AVX512-FCP-NEXT: vporq %xmm2, %xmm4, %xmm18
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm6
+; AVX512-FCP-NEXT: vporq %ymm4, %ymm6, %ymm19
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm4
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512-FCP-NEXT: vporq %xmm4, %xmm8, %xmm20
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
+; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm21
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm22
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm24
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm10
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm10, %ymm23
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm10
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm10, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm1
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm27
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm14[8,u],zero,xmm14[7],zero,xmm14[9,u,11,u],zero,xmm14[10],zero,xmm14[12,u],zero
+; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm3
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm1
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm5
+; AVX512-FCP-NEXT: vpor %xmm1, %xmm5, %xmm5
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm2, %ymm11
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[26],zero,ymm12[28],zero,zero,ymm12[27],zero,ymm12[29],zero,ymm12[31],zero,zero,ymm12[30],zero
+; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm8
-; AVX512-FCP-NEXT: vpermd %ymm8, %ymm11, %ymm11
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512-FCP-NEXT: vpandnq %ymm11, %ymm25, %ymm11
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm13
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm26
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm31 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm11
-; AVX512-FCP-NEXT: vpermd %ymm11, %ymm31, %ymm27
-; AVX512-FCP-NEXT: vpandnq %ymm27, %ymm30, %ymm27
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm27
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm7
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm5
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,ymm14[26],zero,ymm14[28],zero,ymm14[30],zero,zero,ymm14[29],zero,ymm14[31],zero,zero
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm15[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[21],zero,zero,ymm6[20],zero,ymm6[22],zero,ymm6[24],zero,zero,ymm6[23],zero,ymm6[25],zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[26],zero,ymm6[28],zero,zero,ymm6[27],zero,ymm6[29],zero,ymm6[31],zero,zero,ymm6[30],zero
-; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm28, %zmm1, %zmm28
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm7[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm9[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm13[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm5[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm5
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm10
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm3
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm29, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
+; AVX512-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm10
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm2
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm10
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512-FCP-NEXT: vpermd %ymm10, %ymm28, %ymm30
+; AVX512-FCP-NEXT: vpandnq %ymm30, %ymm29, %ymm30
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm9, %zmm3
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3],xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm11
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm21 = zmm21[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400,72056498804555775,72056498804555775,18374967950370078975,18374967950370078975]
-; AVX512-FCP-NEXT: vpternlogq $248, %zmm29, %zmm21, %zmm22
-; AVX512-FCP-NEXT: vpandq %ymm29, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vporq %zmm7, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm22, %zmm7, %zmm1
-; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm31, %zmm8
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm1, %zmm30, %zmm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm10
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm11
+; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm6
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3],xmm6[4],xmm15[4],xmm6[5],xmm15[5],xmm6[6],xmm15[6],xmm6[7],xmm15[7]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm1
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm21[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm22[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vporq %zmm6, %zmm7, %zmm6
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm9, %zmm7
+; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm8
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm8
; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
-; AVX512-FCP-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm1 = mem[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm8 = mem[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm8
-; AVX512-FCP-NEXT: vpternlogq $248, %zmm25, %zmm8, %zmm26
-; AVX512-FCP-NEXT: vpor %ymm4, %ymm9, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm24, %zmm1
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm29, %ymm12, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm20, %zmm2
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm2
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm29, %ymm13, %ymm14
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm14, %zmm1
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm7, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm27
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512-FCP-NEXT: vpermd %zmm5, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm28[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm3[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512-FCP-NEXT: vpermd %zmm11, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 %zmm27, 64(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm26, 192(%r9)
-; AVX512-FCP-NEXT: addq $24, %rsp
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm18[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm20[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm7, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
+; AVX512-FCP-NEXT: vpternlogq $248, %zmm12, %zmm7, %zmm2
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm25[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm23, %zmm6
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm16[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm27, %zmm7
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm11, %zmm9, %zmm1
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
+; AVX512-FCP-NEXT: vpermd %zmm10, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512-FCP-NEXT: vpermd %zmm5, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 64(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i8_stride5_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm3, %ymm0
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm20
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm12, %xmm0
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm28
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm10, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm29
+; AVX512DQ-NEXT: vporq %xmm0, %xmm1, %xmm21
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm15, %ymm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm13
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm13, %ymm1
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm6, %xmm0
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm30
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm7, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm31
+; AVX512DQ-NEXT: vporq %xmm0, %xmm1, %xmm23
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
+; AVX512DQ-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm5, %ymm0
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
+; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm11, %ymm1
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm11, %ymm11
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm26
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512DQ-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm13, %ymm1
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm15, %ymm14
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm14, %zmm25
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512DQ-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm15, %ymm14
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512DQ-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm13, %ymm13
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm27
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm13
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm13, %ymm8
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512DQ-NEXT: vporq %ymm8, %ymm3, %ymm16
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm14, %ymm3
+; AVX512DQ-NEXT: vporq %ymm0, %ymm3, %ymm17
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm8, %ymm15
+; AVX512DQ-NEXT: vporq %ymm0, %ymm15, %ymm18
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm8, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT: vporq %ymm0, %ymm2, %ymm19
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm14, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm5 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512DQ-NEXT: vpshufb %xmm5, %xmm1, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm16
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm14 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm4, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm8
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm8, %ymm4
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm11
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm11, %ymm10
-; AVX512DQ-NEXT: vpor %ymm4, %ymm10, %ymm4
-; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm13
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm13, %xmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm25
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm10, %xmm12
-; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm26
-; AVX512DQ-NEXT: vporq %xmm4, %xmm12, %xmm20
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm22
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm23
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,ymm11[26],zero,ymm11[28],zero,ymm11[30],zero,zero,ymm11[29],zero,ymm11[31],zero,zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm8[19],zero,ymm8[21],zero,zero,ymm8[20],zero,ymm8[22],zero,ymm8[24],zero,zero,ymm8[23],zero
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm24
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm12
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm6
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm6, %ymm1
-; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm19
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm7, %ymm2
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm4, %ymm3
-; AVX512DQ-NEXT: vporq %ymm2, %ymm3, %ymm21
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm5, %xmm0, %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm5, %xmm9
-; AVX512DQ-NEXT: vporq %xmm3, %xmm9, %xmm27
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm9, %xmm15
-; AVX512DQ-NEXT: vporq %xmm0, %xmm15, %xmm29
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = mem[1,1,2,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,1]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm8, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm5, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm9, %xmm3
+; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm4
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm11
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm8, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm13
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm3, %xmm13
+; AVX512DQ-NEXT: vpor %xmm2, %xmm13, %xmm13
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm14 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512DQ-NEXT: vpshufb %ymm14, %ymm11, %ymm2
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm15 = mem[1,1,2,2]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,1,1]
; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm28 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512DQ-NEXT: vpandnq %ymm0, %ymm28, %ymm0
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm15
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm15, %ymm14
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm26
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512DQ-NEXT: vpandnq %ymm0, %ymm30, %ymm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm25
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm8, %ymm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm14 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[19],zero,ymm12[21],zero,zero,ymm12[20],zero,ymm12[22],zero,ymm12[24],zero,zero,ymm12[23],zero
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm12, %ymm12
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm11, %ymm11
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm6, %ymm2
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,ymm6[26],zero,ymm6[28],zero,ymm6[30],zero,zero,ymm6[29],zero,ymm6[31],zero,zero
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3],xmm13[4],xmm10[4],xmm13[5],xmm10[5],xmm13[6],xmm10[6],xmm13[7],xmm10[7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm13 = ymm7[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,ymm4[27],zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30],zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm4
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [4,0,5,5,5,5,0,6,6,6,6,0,7,7,7,7]
-; AVX512DQ-NEXT: vpermd %zmm15, %zmm1, %zmm31
-; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm16
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
-; AVX512DQ-NEXT: vpermi2d %zmm15, %zmm16, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm17, %xmm15
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3],xmm15[4],xmm5[4],xmm15[5],xmm5[5],xmm15[6],xmm5[6],xmm15[7],xmm5[7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm4, %xmm4
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm27, %zmm5, %zmm5
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm15
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3],xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm10, %xmm10
+; AVX512DQ-NEXT: vpandnq %ymm15, %ymm28, %ymm15
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm15, %zmm2
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm14, %ymm15, %ymm14
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,2,1,1,4,6,5,5]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,3,2]
+; AVX512DQ-NEXT: vpandnq %ymm15, %ymm29, %ymm15
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-NEXT: vpshufb %xmm7, %xmm6, %xmm6
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm10, %xmm10
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm9, %xmm9
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm29, %zmm9, %zmm9
-; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm15 = mem[0,0,1,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm15 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm17 = ymm20[0,0,1,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm18 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm15, %zmm18, %zmm17
-; AVX512DQ-NEXT: vpternlogq $248, %zmm28, %zmm17, %zmm26
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm15 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm17 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm20 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400,72056498804555775,72056498804555775,18374967950370078975,18374967950370078975]
-; AVX512DQ-NEXT: vpternlogq $248, %zmm20, %zmm15, %zmm17
-; AVX512DQ-NEXT: vpandq %ymm20, %ymm8, %ymm8
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm8
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm11 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vporq %zmm11, %zmm8, %zmm8
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm17, %zmm11, %zmm8
-; AVX512DQ-NEXT: vpternlogd $184, %zmm8, %zmm30, %zmm31
-; AVX512DQ-NEXT: vpor %ymm2, %ymm14, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm19, %zmm2
-; AVX512DQ-NEXT: vpternlogq $248, %ymm20, %ymm13, %ymm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm18, %zmm0
-; AVX512DQ-NEXT: vpternlogq $248, %ymm20, %ymm12, %ymm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm2
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm3
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm11, %zmm3
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm25
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm9[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm15
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm30 = [4,0,5,5,5,5,0,6,6,6,6,0,7,7,7,7]
+; AVX512DQ-NEXT: vpermd %zmm11, %zmm30, %zmm30
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm31 = [6,6,6,0,7,7,7,7,0,16,16,16,16,0,17,17]
+; AVX512DQ-NEXT: vpermi2d %zmm11, %zmm15, %zmm31
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm9[0],xmm5[1],xmm9[1],xmm5[2],xmm9[2],xmm5[3],xmm9[3],xmm5[4],xmm9[4],xmm5[5],xmm9[5],xmm5[6],xmm9[6],xmm5[7],xmm9[7]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
+; AVX512DQ-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm13, %zmm3, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,0,1,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm20, %zmm5, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm23[0,0,1,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm22, %zmm7, %zmm7
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
+; AVX512DQ-NEXT: vpternlogq $248, %zmm28, %zmm7, %zmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vporq %zmm5, %zmm7, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm9 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm7
+; AVX512DQ-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm30
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm17[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm16, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm19[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm7
+; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm1
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm14
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm31
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm4[0,0,1,1,4,4,5,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512DQ-NEXT: vpermd %zmm16, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-NEXT: vmovdqa64 %zmm25, 64(%r9)
+; AVX512DQ-NEXT: vpermd %zmm15, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqa64 %zmm14, 64(%r9)
; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512DQ-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512DQ-NEXT: vmovdqa64 %zmm31, 256(%r9)
-; AVX512DQ-NEXT: vmovdqa64 %zmm26, 192(%r9)
+; AVX512DQ-NEXT: vmovdqa64 %zmm31, 128(%r9)
+; AVX512DQ-NEXT: vmovdqa64 %zmm30, 256(%r9)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i8_stride5_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $24, %rsp
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm19
-; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm2, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm9, %ymm4
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm12, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm25
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm26
-; AVX512DQ-FCP-NEXT: vpor %xmm2, %xmm4, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm21
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[21],zero,zero,ymm1[20],zero,ymm1[22],zero,ymm1[24],zero,zero,ymm1[23],zero,ymm1[25],zero,zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm22
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm9[27],zero,zero,ymm9[26],zero,ymm9[28],zero,ymm9[30],zero,zero,ymm9[29],zero,ymm9[31],zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[19],zero,ymm7[21],zero,zero,ymm7[20],zero,ymm7[22],zero,ymm7[24],zero,zero,ymm7[23],zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm23
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm14
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm14, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm24
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm20
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm16
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm3, %xmm2
-; AVX512DQ-FCP-NEXT: vporq %xmm1, %xmm2, %xmm28
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm17
-; AVX512DQ-FCP-NEXT: vporq %xmm8, %xmm11, %xmm29
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [1,1,2,2,2,2,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm15[8,u],zero,xmm15[7],zero,xmm15[9,u,11,u],zero,xmm15[10],zero,xmm15[12,u],zero
+; AVX512DQ-FCP-NEXT: vporq %xmm2, %xmm4, %xmm18
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm6
+; AVX512DQ-FCP-NEXT: vporq %ymm4, %ymm6, %ymm19
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512DQ-FCP-NEXT: vporq %xmm4, %xmm8, %xmm20
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
+; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm21
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm22
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm24
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm10
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm10, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm10, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm12
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm27
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm14[8,u],zero,xmm14[7],zero,xmm14[9,u,11,u],zero,xmm14[10],zero,xmm14[12,u],zero
+; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm5
+; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm5, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm2, %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[26],zero,ymm12[28],zero,zero,ymm12[27],zero,ymm12[29],zero,ymm12[31],zero,zero,ymm12[30],zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm8
-; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm11, %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512DQ-FCP-NEXT: vpandnq %ymm11, %ymm25, %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm13
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm26
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm31 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm11
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm31, %ymm27
-; AVX512DQ-FCP-NEXT: vpandnq %ymm27, %ymm30, %ymm27
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm27
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[19],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm5
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,ymm14[26],zero,ymm14[28],zero,ymm14[30],zero,zero,ymm14[29],zero,ymm14[31],zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm15[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[21],zero,zero,ymm6[20],zero,ymm6[22],zero,ymm6[24],zero,zero,ymm6[23],zero,ymm6[25],zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[26],zero,ymm6[28],zero,zero,ymm6[27],zero,ymm6[29],zero,ymm6[31],zero,zero,ymm6[30],zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm28, %zmm1, %zmm28
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm7[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm9[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm13[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm5[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm5
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm3
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm29, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
+; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512DQ-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm10
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm28, %ymm30
+; AVX512DQ-FCP-NEXT: vpandnq %ymm30, %ymm29, %ymm30
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm9, %zmm3
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3],xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm11
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm21 = zmm21[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400,72056498804555775,72056498804555775,18374967950370078975,18374967950370078975]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm29, %zmm21, %zmm22
-; AVX512DQ-FCP-NEXT: vpandq %ymm29, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm22, %zmm7, %zmm1
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm31, %zmm8
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm1, %zmm30, %zmm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm10
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm6
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3],xmm6[4],xmm15[4],xmm6[5],xmm15[5],xmm6[6],xmm15[6],xmm6[7],xmm15[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm1
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm21[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm22[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm6, %zmm7, %zmm6
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm9, %zmm7
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm8
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
-; AVX512DQ-FCP-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm8
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm25, %zmm8, %zmm26
-; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm9, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm24, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm29, %ymm12, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm20, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm29, %ymm13, %ymm14
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm14, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm6
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm7, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm27
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm28[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm3[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm11, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, 64(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, 192(%r9)
-; AVX512DQ-FCP-NEXT: addq $24, %rsp
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm18[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm20[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm7, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm12, %zmm7, %zmm2
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm25[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm23, %zmm6
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm16[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm27, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm11, %zmm9, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm10, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 64(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
; AVX512BW-LABEL: store_i8_stride5_vf64:
; AVX512BW: # %bb.0:
-; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm3
+; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm2
; AVX512BW-NEXT: vmovdqa (%rcx), %ymm0
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512BW-NEXT: vpshufb %ymm8, %ymm0, %ymm2
+; AVX512BW-NEXT: vpshufb %ymm8, %ymm0, %ymm3
; AVX512BW-NEXT: vmovdqa (%rdx), %ymm1
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: movl $693250386, %eax # imm = 0x29522952
; AVX512BW-NEXT: kmovd %eax, %k1
-; AVX512BW-NEXT: vmovdqu8 %ymm4, %ymm2 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512BW-NEXT: vmovdqu8 %ymm4, %ymm3 {%k1}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512BW-NEXT: vmovdqa 32(%rcx), %xmm12
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
; AVX512BW-NEXT: vpshufb %xmm7, %xmm4, %xmm4
; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm10
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm10
; AVX512BW-NEXT: vmovdqa (%rsi), %ymm4
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm15 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
-; AVX512BW-NEXT: vpshufb %ymm15, %ymm4, %ymm2
+; AVX512BW-NEXT: vpshufb %ymm15, %ymm4, %ymm3
; AVX512BW-NEXT: vmovdqa (%rdi), %ymm5
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm5[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: movl $1251232404, %eax # imm = 0x4A944A94
; AVX512BW-NEXT: kmovd %eax, %k5
-; AVX512BW-NEXT: vmovdqu8 %ymm9, %ymm2 {%k5}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512BW-NEXT: vmovdqu8 %ymm9, %ymm3 {%k5}
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512BW-NEXT: vmovdqa 32(%rsi), %xmm13
; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm14
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
; AVX512BW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
; AVX512BW-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512BW-NEXT: kmovq %rax, %k4
-; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k4}
+; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm3 {%k4}
; AVX512BW-NEXT: vmovdqa64 32(%r8), %ymm16
; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm10 = [6,6,6,6,7,7,7,7,16,16,16,16,16,16,17,17]
-; AVX512BW-NEXT: vpermi2d %zmm16, %zmm3, %zmm10
+; AVX512BW-NEXT: vpermi2d %zmm16, %zmm2, %zmm10
; AVX512BW-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
; AVX512BW-NEXT: kmovq %rax, %k2
-; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k2}
+; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm3 {%k2}
; AVX512BW-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
; AVX512BW-NEXT: vpshufb %ymm10, %ymm23, %ymm17
@@ -5584,21 +5491,21 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm21 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
; AVX512BW-NEXT: vpshufb %xmm21, %xmm13, %xmm13
; AVX512BW-NEXT: vpor %xmm12, %xmm13, %xmm12
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm13 = ymm12[0,0,1,1]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm12[0,0,1,1]
; AVX512BW-NEXT: vmovdqa64 32(%rdi), %ymm25
; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm12 = [3,3,3,0,4,4,4,4]
; AVX512BW-NEXT: vpermd %ymm25, %ymm12, %ymm17
; AVX512BW-NEXT: vmovdqa64 32(%rsi), %ymm26
-; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm14 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512BW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512BW-NEXT: kmovd %eax, %k3
-; AVX512BW-NEXT: vpshufb %ymm14, %ymm26, %ymm17 {%k3}
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm17, %zmm13, %zmm13
+; AVX512BW-NEXT: vpshufb %ymm13, %ymm26, %ymm17 {%k3}
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm17, %zmm14, %zmm14
; AVX512BW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512BW-NEXT: kmovq %rax, %k2
-; AVX512BW-NEXT: vmovdqu8 %zmm13, %zmm6 {%k2}
-; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm13 = [3,3,3,3,0,4,4,4]
-; AVX512BW-NEXT: vpermd %ymm16, %ymm13, %ymm17
+; AVX512BW-NEXT: vmovdqu8 %zmm14, %zmm6 {%k2}
+; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,3,3,3,0,4,4,4]
+; AVX512BW-NEXT: vpermd %ymm16, %ymm14, %ymm17
; AVX512BW-NEXT: vpshufd {{.*#+}} xmm18 = mem[1,1,2,2]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,1,1,1]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm17, %zmm18, %zmm17
@@ -5607,32 +5514,28 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqu8 %zmm17, %zmm6 {%k6}
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
; AVX512BW-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpshufb %ymm17, %ymm26, %ymm18
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm27 = ymm18[2,2,3,3]
+; AVX512BW-NEXT: vpshufb %ymm17, %ymm26, %ymm27
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm18 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512BW-NEXT: vpshufb %ymm18, %ymm25, %ymm28
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512BW-NEXT: vporq %ymm27, %ymm28, %ymm27
; AVX512BW-NEXT: vpshufb %ymm15, %ymm26, %ymm15
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm25 = ymm25[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: vmovdqu8 %ymm25, %ymm15 {%k5}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm15, %zmm27, %zmm15
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
; AVX512BW-NEXT: vpshufb %ymm25, %ymm23, %ymm26
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm27 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX512BW-NEXT: # ymm27 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-NEXT: vpshufb %ymm27, %ymm24, %ymm28
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512BW-NEXT: vporq %ymm26, %ymm28, %ymm26
; AVX512BW-NEXT: vpshufb %ymm8, %ymm24, %ymm8
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm23 = ymm23[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm23 = ymm23[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: vmovdqu8 %ymm23, %ymm8 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm26, %zmm8
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k4}
; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm15 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
; AVX512BW-NEXT: vpermd %zmm16, %zmm15, %zmm15
@@ -5661,33 +5564,31 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: kmovq %rax, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm9 {%k1}
; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
-; AVX512BW-NEXT: vpermd %zmm3, %zmm7, %zmm3
+; AVX512BW-NEXT: vpermd %zmm2, %zmm7, %zmm2
; AVX512BW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm3, %zmm9 {%k1}
-; AVX512BW-NEXT: vpshufb %ymm25, %ymm1, %ymm3
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512BW-NEXT: vpshufb %ymm27, %ymm0, %ymm7
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-NEXT: vpor %ymm3, %ymm7, %ymm3
-; AVX512BW-NEXT: vpshufb %ymm10, %ymm1, %ymm1
-; AVX512BW-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX512BW-NEXT: vmovdqu8 %zmm2, %zmm9 {%k1}
+; AVX512BW-NEXT: vpshufb %ymm10, %ymm1, %ymm2
+; AVX512BW-NEXT: vpshufb %ymm11, %ymm0, %ymm7
+; AVX512BW-NEXT: vpor %ymm2, %ymm7, %ymm2
+; AVX512BW-NEXT: vpshufb %ymm25, %ymm1, %ymm1
+; AVX512BW-NEXT: vpshufb %ymm27, %ymm0, %ymm0
; AVX512BW-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT: vpermd %ymm5, %ymm12, %ymm1
-; AVX512BW-NEXT: vpshufb %ymm14, %ymm4, %ymm1 {%k3}
-; AVX512BW-NEXT: vpshufb %ymm17, %ymm4, %ymm3
+; AVX512BW-NEXT: vpshufb %ymm13, %ymm4, %ymm1 {%k3}
+; AVX512BW-NEXT: vpshufb %ymm17, %ymm4, %ymm2
; AVX512BW-NEXT: vpshufb %ymm18, %ymm5, %ymm4
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512BW-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512BW-NEXT: vpor %ymm2, %ymm4, %ymm2
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
; AVX512BW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512BW-NEXT: vpermd %ymm0, %ymm13, %ymm3
+; AVX512BW-NEXT: vpermd %ymm0, %ymm14, %ymm2
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512BW-NEXT: kmovq %rax, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
@@ -5695,215 +5596,206 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa64 %zmm9, (%r9)
; AVX512BW-NEXT: vmovdqa64 %zmm8, 256(%r9)
; AVX512BW-NEXT: vmovdqa64 %zmm6, 192(%r9)
-; AVX512BW-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512BW-NEXT: vmovdqa64 %zmm3, 128(%r9)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BW-FCP-LABEL: store_i8_stride5_vf64:
; AVX512BW-FCP: # %bb.0:
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm5
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm21
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm21, %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm13
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm3
-; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm3, %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm18
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm18, %xmm3
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm20
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm20, %xmm4
-; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm21, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm2
+; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512BW-FCP-NEXT: vpshufb %xmm14, %xmm2, %xmm3
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm4
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm5
+; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm5, %xmm3
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm17
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm17, %xmm3
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm14, %xmm19, %xmm4
-; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm3[0,0,1,1]
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm16
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [3,3,3,0,4,4,4,4]
-; AVX512BW-FCP-NEXT: vpermd %ymm16, %ymm3, %ymm22
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm13
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm17 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
+; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm3, %xmm9
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm16
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm20 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm5, %xmm10
+; AVX512BW-FCP-NEXT: vpor %xmm9, %xmm10, %xmm9
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm9[0,0,1,1]
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm19
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [3,3,3,0,4,4,4,4]
+; AVX512BW-FCP-NEXT: vpermd %ymm19, %ymm9, %ymm22
; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm23
-; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm4 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm10 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512BW-FCP-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm23, %ymm22 {%k1}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm15, %zmm15
+; AVX512BW-FCP-NEXT: vpshufb %ymm10, %ymm23, %ymm22 {%k1}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm18, %zmm18
; AVX512BW-FCP-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm0 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm24
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [1,1,2,2,2,2,2,2,27,27,27,27,0,28,28,28]
-; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm24, %zmm15
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm1 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm22
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [1,1,2,2,2,2,2,2,27,27,27,27,0,28,28,28]
+; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm22, %zmm18
; AVX512BW-FCP-NEXT: movabsq $4760450083537948804, %rax # imm = 0x4210842108421084
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm0 {%k3}
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm23[0,1,2,3],mem[4,5,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
-; AVX512BW-FCP-NEXT: vpshufb %zmm15, %zmm22, %zmm22
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm16[0,1,2,3],mem[4,5,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512BW-FCP-NEXT: vpshufb %zmm16, %zmm23, %zmm23
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm1 {%k3}
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],mem[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
+; AVX512BW-FCP-NEXT: vpshufb %zmm18, %zmm23, %zmm23
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm19[0,1,2,3],mem[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
+; AVX512BW-FCP-NEXT: vpshufb %zmm19, %zmm24, %zmm24
+; AVX512BW-FCP-NEXT: vporq %zmm23, %zmm24, %zmm23
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm23 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512BW-FCP-NEXT: vporq %zmm22, %zmm23, %zmm23
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm21[0,1,2,3],mem[4,5,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
-; AVX512BW-FCP-NEXT: vpshufb %zmm21, %zmm22, %zmm22
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm25 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm13[0,1,2,3],mem[4,5,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %zmm22, %zmm13, %zmm13
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[2,2,3,3,6,6,7,7]
-; AVX512BW-FCP-NEXT: vporq %zmm25, %zmm13, %zmm13
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[0,1,2,3],mem[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
+; AVX512BW-FCP-NEXT: vpshufb %zmm24, %zmm21, %zmm21
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,2,3],mem[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %zmm25, %zmm8, %zmm8
+; AVX512BW-FCP-NEXT: vporq %zmm21, %zmm8, %zmm8
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
; AVX512BW-FCP-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm13 {%k3}
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm23 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512BW-FCP-NEXT: vpermi2d %zmm5, %zmm24, %zmm23
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm8 {%k3}
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm21 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm22, %zmm21
; AVX512BW-FCP-NEXT: movabsq $-8925843906633654008, %rax # imm = 0x8421084210842108
; AVX512BW-FCP-NEXT: kmovq %rax, %k4
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm13 {%k4}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm23
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm24 = ymm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm23[27],zero,zero,ymm23[26],zero,ymm23[28],zero,ymm23[30],zero,zero,ymm23[29],zero,ymm23[31],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm24[2,2,3,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm26 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm24[27],zero,zero,ymm24[26],zero,ymm24[28],zero,ymm24[30],zero,zero,ymm24[29],zero,ymm24[31],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
-; AVX512BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm18[0],xmm20[0],xmm18[1],xmm20[1],xmm18[2],xmm20[2],xmm18[3],xmm20[3],xmm18[4],xmm20[4],xmm18[5],xmm20[5],xmm18[6],xmm20[6],xmm18[7],xmm20[7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm26 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512BW-FCP-NEXT: vpshufb %xmm26, %xmm18, %xmm18
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,0,1,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm25, %zmm25
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm18
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm20 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm18[26],zero,ymm18[28],zero,zero,zero,zero,ymm18[29],zero,ymm18[31],zero,zero,ymm18[30]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm27 = ymm20[2,2,3,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm20
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm28 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm20[26],zero,ymm20[28],zero,zero,ymm20[27],zero,ymm20[29],zero,ymm20[31],zero,zero,ymm20[30],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
-; AVX512BW-FCP-NEXT: vporq %ymm27, %ymm28, %ymm27
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm17, %xmm17
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[0,0,1,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm27, %zmm17
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm25, %zmm17 {%k3}
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [6,6,6,6,7,7,7,7,8,8,8,8,8,8,9,9]
-; AVX512BW-FCP-NEXT: vpermd %zmm5, %zmm25, %zmm5
-; AVX512BW-FCP-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm17 {%k3}
-; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm5
-; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm8
-; AVX512BW-FCP-NEXT: vpor %xmm5, %xmm8, %xmm5
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm26, %xmm6, %xmm6
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm6, %zmm5
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm6
-; AVX512BW-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm7
-; AVX512BW-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3],xmm9[4],xmm11[4],xmm9[5],xmm11[5],xmm9[6],xmm11[6],xmm9[7],xmm11[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm7, %xmm7
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[0,0,1,1,4,4,5,5]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm21, %zmm8 {%k4}
+; AVX512BW-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm14
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm12, %xmm15
+; AVX512BW-FCP-NEXT: vpor %xmm14, %xmm15, %xmm14
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3],xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm14, %zmm11, %zmm11
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm14 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm13, %xmm11
+; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm16, %xmm15
+; AVX512BW-FCP-NEXT: vpor %xmm11, %xmm15, %xmm11
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm16[0],xmm13[1],xmm16[1],xmm13[2],xmm16[2],xmm13[3],xmm16[3],xmm13[4],xmm16[4],xmm13[5],xmm16[5],xmm13[6],xmm16[6],xmm13[7],xmm16[7]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm13, %xmm13
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm11, %zmm13, %zmm11
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,0,1,1,4,4,5,5]
; AVX512BW-FCP-NEXT: movabsq $-4165393823095705204, %rax # imm = 0xC6318C6318C6318C
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm6 {%k3}
-; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512BW-FCP-NEXT: vpermd %zmm5, %zmm7, %zmm7
+; AVX512BW-FCP-NEXT: kmovq %rax, %k4
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm11 {%k4}
+; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
+; AVX512BW-FCP-NEXT: vpermd %zmm13, %zmm14, %zmm14
; AVX512BW-FCP-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k3}
-; AVX512BW-FCP-NEXT: vpshufb %ymm21, %ymm24, %ymm7
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm22, %ymm23, %ymm8
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm24, %ymm1
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm23, %ymm2
-; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
-; AVX512BW-FCP-NEXT: vpshufb %ymm15, %ymm18, %ymm2
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm16, %ymm20, %ymm7
+; AVX512BW-FCP-NEXT: kmovq %rax, %k4
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm11 {%k4}
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm14, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm16
+; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm16, %ymm7
+; AVX512BW-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512BW-FCP-NEXT: vpshufb %ymm24, %ymm14, %ymm7
+; AVX512BW-FCP-NEXT: vpshufb %ymm25, %ymm16, %ymm17
+; AVX512BW-FCP-NEXT: vporq %ymm7, %ymm17, %ymm7
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm7, %ymm2
-; AVX512BW-FCP-NEXT: vpermd %ymm20, %ymm3, %ymm3
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm18, %ymm3 {%k1}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k2}
-; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
-; AVX512BW-FCP-NEXT: vpermd %zmm5, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512BW-FCP-NEXT: vpshufb %ymm18, %ymm7, %ymm17
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm18
+; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm18, %ymm19
+; AVX512BW-FCP-NEXT: vporq %ymm17, %ymm19, %ymm17
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,3,3]
+; AVX512BW-FCP-NEXT: vpermd %ymm18, %ymm9, %ymm9
+; AVX512BW-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm9 {%k1}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm9
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k2}
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
+; AVX512BW-FCP-NEXT: vpermd %zmm13, %zmm6, %zmm6
; AVX512BW-FCP-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%r9)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, (%r9)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 128(%r9)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 256(%r9)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%r9)
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k1}
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm16[27],zero,zero,ymm16[26],zero,ymm16[28],zero,ymm16[30],zero,zero,ymm16[29],zero,ymm16[31],zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,ymm14[26],zero,ymm14[28],zero,ymm14[30],zero,zero,ymm14[29],zero,ymm14[31],zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
+; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [2,2,3,3,8,8,9,9]
+; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm4
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm7[26],zero,ymm7[28],zero,zero,zero,zero,ymm7[29],zero,ymm7[31],zero,zero,ymm7[30]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm18[26],zero,ymm18[28],zero,zero,ymm18[27],zero,ymm18[29],zero,ymm18[31],zero,zero,ymm18[30],zero
+; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm5, %ymm3
+; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k3}
+; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,6,7,7,7,7,8,8,8,8,8,8,9,9]
+; AVX512BW-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
+; AVX512BW-FCP-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%r9)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 64(%r9)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, (%r9)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%r9)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
; AVX512DQ-BW-LABEL: store_i8_stride5_vf64:
; AVX512DQ-BW: # %bb.0:
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm3
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm2
; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm0
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm0, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm0, %ymm3
; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm1
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: movl $693250386, %eax # imm = 0x29522952
; AVX512DQ-BW-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm4, %ymm2 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm4, %ymm3 {%k1}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512DQ-BW-NEXT: vmovdqa 32(%rcx), %xmm12
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm4, %xmm4
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm10
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm10
; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm4
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm15 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
-; AVX512DQ-BW-NEXT: vpshufb %ymm15, %ymm4, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %ymm15, %ymm4, %ymm3
; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm5
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm5[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: movl $1251232404, %eax # imm = 0x4A944A94
; AVX512DQ-BW-NEXT: kmovd %eax, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm9, %ymm2 {%k5}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm9, %ymm3 {%k5}
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %xmm13
; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm14
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
; AVX512DQ-BW-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512DQ-BW-NEXT: kmovq %rax, %k4
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k4}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm3 {%k4}
; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %ymm16
; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm10 = [6,6,6,6,7,7,7,7,16,16,16,16,16,16,17,17]
-; AVX512DQ-BW-NEXT: vpermi2d %zmm16, %zmm3, %zmm10
+; AVX512DQ-BW-NEXT: vpermi2d %zmm16, %zmm2, %zmm10
; AVX512DQ-BW-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
; AVX512DQ-BW-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm3 {%k2}
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %ymm23
; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
; AVX512DQ-BW-NEXT: vpshufb %ymm10, %ymm23, %ymm17
@@ -5923,21 +5815,21 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm21 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm13, %xmm13
; AVX512DQ-BW-NEXT: vpor %xmm12, %xmm13, %xmm12
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm13 = ymm12[0,0,1,1]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm12[0,0,1,1]
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %ymm25
; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm12 = [3,3,3,0,4,4,4,4]
; AVX512DQ-BW-NEXT: vpermd %ymm25, %ymm12, %ymm17
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %ymm26
-; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm14 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512DQ-BW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512DQ-BW-NEXT: kmovd %eax, %k3
-; AVX512DQ-BW-NEXT: vpshufb %ymm14, %ymm26, %ymm17 {%k3}
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm17, %zmm13, %zmm13
+; AVX512DQ-BW-NEXT: vpshufb %ymm13, %ymm26, %ymm17 {%k3}
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm17, %zmm14, %zmm14
; AVX512DQ-BW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512DQ-BW-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm13, %zmm6 {%k2}
-; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm13 = [3,3,3,3,0,4,4,4]
-; AVX512DQ-BW-NEXT: vpermd %ymm16, %ymm13, %ymm17
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm14, %zmm6 {%k2}
+; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm14 = [3,3,3,3,0,4,4,4]
+; AVX512DQ-BW-NEXT: vpermd %ymm16, %ymm14, %ymm17
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm18 = mem[1,1,2,2]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,1,1,1]
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm17, %zmm18, %zmm17
@@ -5946,32 +5838,28 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm17, %zmm6 {%k6}
; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
; AVX512DQ-BW-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm26, %ymm18
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm27 = ymm18[2,2,3,3]
+; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm26, %ymm27
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm18 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512DQ-BW-NEXT: vpshufb %ymm18, %ymm25, %ymm28
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512DQ-BW-NEXT: vporq %ymm27, %ymm28, %ymm27
; AVX512DQ-BW-NEXT: vpshufb %ymm15, %ymm26, %ymm15
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm25 = ymm25[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm25, %ymm15 {%k5}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm15, %zmm27, %zmm15
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm23, %ymm26
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm27 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX512DQ-BW-NEXT: # ymm27 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm24, %ymm28
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512DQ-BW-NEXT: vporq %ymm26, %ymm28, %ymm26
; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm24, %ymm8
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm23 = ymm23[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm23 = ymm23[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm23, %ymm8 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm26, %zmm8
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k4}
; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm15 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
; AVX512DQ-BW-NEXT: vpermd %zmm16, %zmm15, %zmm15
@@ -6000,33 +5888,31 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm9 {%k1}
; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
-; AVX512DQ-BW-NEXT: vpermd %zmm3, %zmm7, %zmm3
+; AVX512DQ-BW-NEXT: vpermd %zmm2, %zmm7, %zmm2
; AVX512DQ-BW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm3, %zmm9 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm1, %ymm3
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm0, %ymm7
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpor %ymm3, %ymm7, %ymm3
-; AVX512DQ-BW-NEXT: vpshufb %ymm10, %ymm1, %ymm1
-; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm2, %zmm9 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb %ymm10, %ymm1, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm0, %ymm7
+; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm7, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm1, %ymm1
+; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm0, %ymm0
; AVX512DQ-BW-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT: vpermd %ymm5, %ymm12, %ymm1
-; AVX512DQ-BW-NEXT: vpshufb %ymm14, %ymm4, %ymm1 {%k3}
-; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm4, %ymm3
+; AVX512DQ-BW-NEXT: vpshufb %ymm13, %ymm4, %ymm1 {%k3}
+; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm4, %ymm2
; AVX512DQ-BW-NEXT: vpshufb %ymm18, %ymm5, %ymm4
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512DQ-BW-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm4, %ymm2
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-BW-NEXT: vpermd %ymm0, %ymm13, %ymm3
+; AVX512DQ-BW-NEXT: vpermd %ymm0, %ymm14, %ymm2
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
@@ -6034,166 +5920,157 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, (%r9)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 256(%r9)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 192(%r9)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 128(%r9)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
; AVX512DQ-BW-FCP-LABEL: store_i8_stride5_vf64:
; AVX512DQ-BW-FCP: # %bb.0:
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm21
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm21, %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm13
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm3, %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm18
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm18, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm20
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm20, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm21, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm14, %xmm2, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm5
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm5, %xmm3
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm17
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm17, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm14, %xmm19, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm3[0,0,1,1]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm16
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [3,3,3,0,4,4,4,4]
-; AVX512DQ-BW-FCP-NEXT: vpermd %ymm16, %ymm3, %ymm22
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm13
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm17 = [8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm3, %xmm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm16
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm20 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm5, %xmm10
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm9, %xmm10, %xmm9
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm9[0,0,1,1]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm19
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [3,3,3,0,4,4,4,4]
+; AVX512DQ-BW-FCP-NEXT: vpermd %ymm19, %ymm9, %ymm22
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm23
-; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm4 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm10 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512DQ-BW-FCP-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm23, %ymm22 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm15, %zmm15
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm10, %ymm23, %ymm22 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm18, %zmm18
; AVX512DQ-BW-FCP-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm0 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm24
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [1,1,2,2,2,2,2,2,27,27,27,27,0,28,28,28]
-; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm24, %zmm15
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm1 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm22
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [1,1,2,2,2,2,2,2,27,27,27,27,0,28,28,28]
+; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm22, %zmm18
; AVX512DQ-BW-FCP-NEXT: movabsq $4760450083537948804, %rax # imm = 0x4210842108421084
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm0 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm23[0,1,2,3],mem[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm15, %zmm22, %zmm22
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm16[0,1,2,3],mem[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm16, %zmm23, %zmm23
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm1 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],mem[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,128,128,26,128,28,128,128,128,128,29,128,31,128,128,30]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm18, %zmm23, %zmm23
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm19[0,1,2,3],mem[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm19, %zmm24, %zmm24
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm23, %zmm24, %zmm23
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm23 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vporq %zmm22, %zmm23, %zmm23
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm21[0,1,2,3],mem[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm21, %zmm22, %zmm22
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm25 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm13[0,1,2,3],mem[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm22, %zmm13, %zmm13
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vporq %zmm25, %zmm13, %zmm13
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[0,1,2,3],mem[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm24, %zmm21, %zmm21
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,2,3],mem[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm25, %zmm8, %zmm8
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm21, %zmm8, %zmm8
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm13 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm23 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm5, %zmm24, %zmm23
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm8 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm21 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm22, %zmm21
; AVX512DQ-BW-FCP-NEXT: movabsq $-8925843906633654008, %rax # imm = 0x8421084210842108
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k4
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm13 {%k4}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm23
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm24 = ymm23[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm23[27],zero,zero,ymm23[26],zero,ymm23[28],zero,ymm23[30],zero,zero,ymm23[29],zero,ymm23[31],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm24[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm26 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm24[27],zero,zero,ymm24[26],zero,ymm24[28],zero,ymm24[30],zero,zero,ymm24[29],zero,ymm24[31],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm18[0],xmm20[0],xmm18[1],xmm20[1],xmm18[2],xmm20[2],xmm18[3],xmm20[3],xmm18[4],xmm20[4],xmm18[5],xmm20[5],xmm18[6],xmm20[6],xmm18[7],xmm20[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm26 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm26, %xmm18, %xmm18
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,0,1,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm25, %zmm25
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm20 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm18[26],zero,ymm18[28],zero,zero,zero,zero,ymm18[29],zero,ymm18[31],zero,zero,ymm18[30]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm27 = ymm20[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm20
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm28 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm20[26],zero,ymm20[28],zero,zero,ymm20[27],zero,ymm20[29],zero,ymm20[31],zero,zero,ymm20[30],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm27, %ymm28, %ymm27
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm17, %xmm17
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[0,0,1,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm27, %zmm17
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm25, %zmm17 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [6,6,6,6,7,7,7,7,8,8,8,8,8,8,9,9]
-; AVX512DQ-BW-FCP-NEXT: vpermd %zmm5, %zmm25, %zmm5
-; AVX512DQ-BW-FCP-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm17 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm8
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm5, %xmm8, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm26, %xmm6, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm6, %zmm5
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm7
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3],xmm9[4],xmm11[4],xmm9[5],xmm11[5],xmm9[6],xmm11[6],xmm9[7],xmm11[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm7, %xmm7
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm7, %zmm6
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm21, %zmm8 {%k4}
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm14
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm12, %xmm15
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm14, %xmm15, %xmm14
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3],xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm14, %zmm11, %zmm11
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm14 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm13, %xmm11
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm16, %xmm15
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm11, %xmm15, %xmm11
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm16[0],xmm13[1],xmm16[1],xmm13[2],xmm16[2],xmm13[3],xmm16[3],xmm13[4],xmm16[4],xmm13[5],xmm16[5],xmm13[6],xmm16[6],xmm13[7],xmm16[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm13, %xmm13
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm11, %zmm13, %zmm11
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,0,1,1,4,4,5,5]
; AVX512DQ-BW-FCP-NEXT: movabsq $-4165393823095705204, %rax # imm = 0xC6318C6318C6318C
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm6 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
-; AVX512DQ-BW-FCP-NEXT: vpermd %zmm5, %zmm7, %zmm7
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm11 {%k4}
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,0,0,1,1,9,9,10,10,10,10,10,10]
+; AVX512DQ-BW-FCP-NEXT: vpermd %zmm13, %zmm14, %zmm14
; AVX512DQ-BW-FCP-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm21, %ymm24, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm22, %ymm23, %ymm8
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm24, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm23, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm15, %ymm18, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm16, %ymm20, %ymm7
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm11 {%k4}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm14, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm16
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm16, %ymm7
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm24, %ymm14, %ymm7
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm25, %ymm16, %ymm17
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm7, %ymm17, %ymm7
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm7, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpermd %ymm20, %ymm3, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm18, %ymm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
-; AVX512DQ-BW-FCP-NEXT: vpermd %zmm5, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm18, %ymm7, %ymm17
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm18
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm18, %ymm19
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm17, %ymm19, %ymm17
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,3,3]
+; AVX512DQ-BW-FCP-NEXT: vpermd %ymm18, %ymm9, %ymm9
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm9 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [3,3,3,3,0,4,4,4,12,14,13,13,13,13,12,14]
+; AVX512DQ-BW-FCP-NEXT: vpermd %zmm13, %zmm6, %zmm6
; AVX512DQ-BW-FCP-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%r9)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, (%r9)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 128(%r9)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 256(%r9)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm16[27],zero,zero,ymm16[26],zero,ymm16[28],zero,ymm16[30],zero,zero,ymm16[29],zero,ymm16[31],zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,ymm14[26],zero,ymm14[28],zero,ymm14[30],zero,zero,ymm14[29],zero,ymm14[31],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [2,2,3,3,8,8,9,9]
+; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,ymm7[26],zero,ymm7[28],zero,zero,zero,zero,ymm7[29],zero,ymm7[31],zero,zero,ymm7[30]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm18[26],zero,ymm18[28],zero,zero,ymm18[27],zero,ymm18[29],zero,ymm18[31],zero,zero,ymm18[30],zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm5, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,6,7,7,7,7,8,8,8,8,8,8,9,9]
+; AVX512DQ-BW-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
+; AVX512DQ-BW-FCP-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 64(%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, (%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%r9)
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
index 0495e240ba968a..8b6ba51506ab79 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
@@ -3689,10 +3689,9 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
; AVX2-NEXT: vpblendvb %ymm9, %ymm0, %ymm8, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[27],zero,ymm7[27,28,29,30],zero,ymm7[28],zero,ymm7[26,27,30,31],zero,ymm7[29]
-; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
; AVX2-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u]
; AVX2-NEXT: vpblendvb %ymm9, %ymm0, %ymm8, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
@@ -3772,18 +3771,16 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vpblendvb %ymm7, %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[18],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22],zero,ymm5[20]
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpor %ymm1, %ymm7, %ymm1
+; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
; AVX2-NEXT: vpblendvb %ymm7, %ymm1, %ymm0, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-NEXT: vpshufb {{.*#+}} ymm7 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm12[18],zero,zero,zero,zero,ymm12[21],zero,ymm12[19],zero,zero,zero,zero,ymm12[22],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpor %ymm1, %ymm7, %ymm1
+; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-NEXT: vpshuflw {{.*#+}} ymm7 = ymm11[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX2-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,1,1,3,4,5,5,7]
@@ -3793,23 +3790,20 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
; AVX2-NEXT: vpblendvb %ymm7, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[23],zero,ymm4[27,20,21,26],zero,ymm4[24],zero,ymm4[26,27,26,27],zero,ymm4[25]
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpor %ymm1, %ymm7, %ymm1
+; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm7 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm5[25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27]
-; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
; AVX2-NEXT: vpor %ymm7, %ymm8, %ymm7
+; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm8 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-NEXT: vpblendvb %ymm8, %ymm1, %ymm7, %ymm1
; AVX2-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm12[25],zero,ymm12[23],zero,zero,zero,zero,ymm12[26],zero,ymm12[24],zero,zero,zero
; AVX2-NEXT: vmovdqa %ymm12, %ymm13
-; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
; AVX2-NEXT: vpor %ymm7, %ymm8, %ymm7
+; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm8 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX2-NEXT: vmovdqa %ymm11, %ymm12
; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
@@ -3919,22 +3913,19 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255]
; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm9, %ymm10, %ymm9
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[17,18,19,30],zero,ymm0[28],zero,ymm0[28,29,30,31],zero,ymm0[29],zero,ymm0[31]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm10, %ymm11, %ymm10
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm11, %ymm12, %ymm11
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm10, %ymm11, %ymm10
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,ymm6[27,28,29,30],zero,ymm6[28],zero,ymm6[26,27,30,31],zero,ymm6[29]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm8[27],zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm11, %ymm12, %ymm11
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u]
; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm10, %ymm11, %ymm10
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
@@ -3942,22 +3933,19 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0]
; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm10, %ymm11, %ymm10
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[23],zero,ymm1[27,20,21,26],zero,ymm1[24],zero,ymm1[26,27,26,27],zero,ymm1[25]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm11, %ymm12, %ymm11
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[25],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero,ymm0[27]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-FP-NEXT: vpblendvb %ymm13, %ymm11, %ymm12, %ymm11
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[25],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm14 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
@@ -3965,22 +3953,19 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm13 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
; AVX2-FP-NEXT: vpblendvb %ymm13, %ymm11, %ymm12, %ymm11
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22],zero,ymm2[20]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20],zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm13, %ymm14, %ymm13
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm14 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm12, %ymm13, %ymm12
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm8[20],zero,ymm8[18],zero,zero,zero,zero,ymm8[21],zero,ymm8[19],zero,zero,zero,zero,ymm8[22]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm13, %ymm14, %ymm13
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,18,19,18,19,20,21,18,19,20,21,28,29,30,31]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
@@ -4089,46 +4074,40 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm11, %ymm9, %ymm10, %ymm9
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22],zero,ymm2[20]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm10, %ymm11, %ymm10
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20],zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm11, %ymm12, %ymm11
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm10, %ymm11, %ymm10
; AVX2-FCP-NEXT: vpshuflw {{.*#+}} ymm11 = ymm7[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [4,5,4,5,5,7,4,5]
; AVX2-FCP-NEXT: vpermd %ymm11, %ymm12, %ymm11
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm8[20],zero,ymm8[18],zero,zero,zero,zero,ymm8[21],zero,ymm8[19],zero,zero,zero,zero,ymm8[22]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm12, %ymm11, %ymm11
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm10, %ymm11, %ymm10
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[17,18,19,30],zero,ymm0[28],zero,ymm0[28,29,30,31],zero,ymm0[29],zero,ymm0[31]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm11, %ymm12, %ymm11
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm11, %ymm12, %ymm11
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,ymm6[27,28,29,30],zero,ymm6[28],zero,ymm6[26,27,30,31],zero,ymm6[29]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm8[27],zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u,255,255,255,255,0,0,u]
; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm11, %ymm12, %ymm11
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
@@ -4136,22 +4115,19 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0]
; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm11, %ymm12, %ymm11
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[23],zero,ymm1[27,20,21,26],zero,ymm1[24],zero,ymm1[26,27,26,27],zero,ymm1[25]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[25],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero,ymm0[27]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm13, %ymm14, %ymm13
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-FCP-NEXT: vpblendvb %ymm14, %ymm12, %ymm13, %ymm12
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[25],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm13, %ymm14, %ymm13
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
@@ -4189,153 +4165,138 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512: # %bb.0:
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512-NEXT: vmovdqa (%r8), %ymm5
-; AVX512-NEXT: vmovdqa (%r9), %ymm6
-; AVX512-NEXT: vmovdqa (%r10), %ymm4
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm8 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm9 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512-NEXT: vpermi2d %zmm7, %zmm8, %zmm9
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero
+; AVX512-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512-NEXT: vmovdqa (%rdx), %ymm5
+; AVX512-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512-NEXT: vmovdqa (%r8), %ymm1
+; AVX512-NEXT: vmovdqa (%r9), %ymm2
+; AVX512-NEXT: vmovdqa (%r10), %ymm0
+; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero,ymm3[25]
; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm5[18],zero,ymm5[20,21,20,21],zero,ymm5[19],zero,ymm5[19,20,21,22],zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm5[23],zero,ymm5[23,24,25,26],zero,ymm5[24],zero,ymm5[30,31]
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm4[23,u,u,u],zero,ymm4[26],zero,ymm4[24,u,u,u],zero,ymm4[27],zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
; AVX512-NEXT: vporq %zmm7, %zmm8, %zmm7
-; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
-; AVX512-NEXT: vpandq %ymm16, %ymm8, %ymm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm11[18,19,20,21],zero,ymm11[19],zero,ymm11[25,26,27,22],zero,ymm11[20],zero
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm3[19],zero,ymm3[21,20,21,22],zero,ymm3[20],zero,ymm3[22,23]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
; AVX512-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm9, %zmm8, %zmm9
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512-NEXT: vpandq %ymm17, %ymm8, %ymm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm11[23],zero,ymm11[21,22,23,26],zero,ymm11[24],zero,ymm11[28,29,26,27]
-; AVX512-NEXT: vmovdqa64 %ymm11, %ymm20
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero
-; AVX512-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
-; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm10, %zmm8, %zmm8
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
-; AVX512-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm11[u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6,u,u,u],zero
-; AVX512-NEXT: vmovdqa (%rdi), %xmm12
-; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm12[u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u,u,9]
-; AVX512-NEXT: vpor %xmm7, %xmm9, %xmm7
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512-NEXT: vinserti32x4 $2, %xmm7, %zmm9, %zmm7
-; AVX512-NEXT: vpermq {{.*#+}} zmm10 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vmovdqa (%rcx), %xmm7
-; AVX512-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[u,u,u],zero,xmm7[7],zero,xmm7[5,u,u,u],zero,xmm7[8],zero,xmm7[6,u,u]
-; AVX512-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512-NEXT: vpshufb {{.*#+}} xmm14 = xmm9[u,u,u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u]
-; AVX512-NEXT: vpor %xmm13, %xmm14, %xmm13
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512-NEXT: vinserti32x4 $2, %xmm13, %zmm14, %zmm13
-; AVX512-NEXT: vpermq {{.*#+}} zmm18 = zmm13[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm18
-; AVX512-NEXT: vmovdqa (%r9), %xmm13
-; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = zero,xmm13[4,u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6]
-; AVX512-NEXT: vmovdqa (%r8), %xmm14
-; AVX512-NEXT: vpshufb {{.*#+}} xmm15 = xmm14[4],zero,xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero
-; AVX512-NEXT: vpor %xmm10, %xmm15, %xmm10
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10
-; AVX512-NEXT: vpermq {{.*#+}} zmm19 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vmovdqa (%r10), %xmm15
-; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = xmm15[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm15[1,1,0,0,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
-; AVX512-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,0,1,0,4,4,5,4]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm10
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm10
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14,u,u],zero,zero,zero,zero,ymm1[15,u,u],zero,zero,zero,zero,ymm1[16,u,u],zero,zero,zero,zero,ymm1[17,u,u],zero,zero,zero,zero,ymm1[18]
+; AVX512-NEXT: vporq %zmm8, %zmm9, %zmm8
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
+; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm10 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512-NEXT: vpermi2d %zmm7, %zmm9, %zmm10
+; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
; AVX512-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
+; AVX512-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm7
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm7
+; AVX512-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm9[u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u,u],zero
+; AVX512-NEXT: vmovdqa (%rdi), %xmm10
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm10[u,7],zero,xmm10[5],zero,xmm10[u,u,u,8],zero,xmm10[6],zero,xmm10[u,u,u,9]
+; AVX512-NEXT: vpor %xmm8, %xmm11, %xmm8
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-NEXT: vinserti32x4 $2, %xmm8, %zmm11, %zmm8
+; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm14[u,u,u],zero,xmm14[7],zero,xmm14[5,u,u,u],zero,xmm14[8],zero,xmm14[6,u,u]
+; AVX512-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512-NEXT: vpshufb {{.*#+}} xmm12 = xmm15[u,u,u,7],zero,xmm15[5],zero,xmm15[u,u,u,8],zero,xmm15[6],zero,xmm15[u,u]
+; AVX512-NEXT: vpor %xmm11, %xmm12, %xmm11
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-NEXT: vinserti32x4 $2, %xmm11, %zmm12, %zmm11
+; AVX512-NEXT: vpermq {{.*#+}} zmm16 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm16
+; AVX512-NEXT: vmovdqa (%r9), %xmm11
+; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm11[4,u,u,u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6]
+; AVX512-NEXT: vmovdqa (%r8), %xmm12
+; AVX512-NEXT: vpshufb {{.*#+}} xmm13 = xmm12[4],zero,xmm12[u,u,u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero
+; AVX512-NEXT: vpor %xmm8, %xmm13, %xmm8
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-NEXT: vinserti32x4 $2, %xmm8, %zmm13, %zmm8
+; AVX512-NEXT: vpermq {{.*#+}} zmm17 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vmovdqa (%r10), %xmm13
+; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm13[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm13[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm8
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm8
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm4[14,u,u],zero,zero,zero,zero,ymm4[15,u,u],zero,zero,zero,zero,ymm4[16,u,u],zero,zero,zero,zero,ymm4[17,u,u],zero,zero,zero,zero,ymm4[18]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,14],zero,ymm3[u,u,0,1,14,15],zero,ymm3[u,u,13,2,3,16],zero,ymm3[u,u,28,29,16,17],zero,ymm3[u,u,19,28,29,18],zero
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm15[8],xmm14[9],xmm15[9],xmm14[10],xmm15[10],xmm14[11],xmm15[11],xmm14[12],xmm15[12],xmm14[13],xmm15[13],xmm14[14],xmm15[14],xmm14[15],xmm15[15]
; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u,u,u]
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm2
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,14],zero,ymm2[u,u,u,u,u,15],zero,ymm2[u,u,u,u,u,16],zero,ymm2[u,u,u,u,u,17],zero,ymm2[u,u,u,u,u]
-; AVX512-NEXT: vpor %ymm1, %ymm7, %ymm1
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm14 = ymm5[u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u,u,u]
+; AVX512-NEXT: vpor %ymm1, %ymm14, %ymm1
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm5[u,u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[13,u,u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u]
-; AVX512-NEXT: vpor %ymm0, %ymm7, %ymm0
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm15[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm7
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm4[13,u,u,u,u],zero,zero,ymm4[14,u,u,u,u],zero,zero,ymm4[15,u,u,u,u],zero,zero,ymm4[16,u,u,u,u],zero,zero,ymm4[17,u,u]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm14
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u,u,u,u,u,14],zero,ymm14[u,u,u,u,u,15],zero,ymm14[u,u,u,u,u,16],zero,ymm14[u,u,u,u,u,17],zero,ymm14[u,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
+; AVX512-NEXT: vpor %ymm0, %ymm9, %ymm0
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm13[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm11
+; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm11[13,u,u,u,u],zero,zero,ymm11[14,u,u,u,u],zero,zero,ymm11[15,u,u,u,u],zero,zero,ymm11[16,u,u,u,u],zero,zero,ymm11[17,u,u]
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm9
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm6[30],zero,ymm6[28,u,u,u],zero,ymm6[31],zero,ymm6[29,u]
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-NEXT: vpternlogq $248, %ymm16, %ymm0, %ymm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-NEXT: vpternlogq $248, %ymm17, %ymm1, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm3
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
-; AVX512-NEXT: vmovdqa %ymm3, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm10, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm2
+; AVX512-NEXT: vmovdqa %ymm2, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm9, 64(%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -4343,12 +4304,12 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP: # %bb.0:
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm3
; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm4
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm5
; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm6
; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm2
; AVX512-FCP-NEXT: vmovdqa64 (%r10), %ymm17
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm8
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm8[u],zero,xmm8[7],zero,xmm8[5,u,u,u],zero,xmm8[8],zero,xmm8[6,u,u,u],zero
@@ -4388,7 +4349,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm7
; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm4[14,u,u],zero,zero,zero,zero,ymm4[15,u,u],zero,zero,zero,zero,ymm4[16,u,u],zero,zero,zero,zero,ymm4[17,u,u],zero,zero,zero,zero,ymm4[18]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm3[0,1,14],zero,ymm3[u,u,0,1,14,15],zero,ymm3[u,u,13,2,3,16],zero,ymm3[u,u,28,29,16,17],zero,ymm3[u,u,19,28,29,18],zero
; AVX512-FCP-NEXT: vpor %ymm0, %ymm15, %ymm0
; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
@@ -4403,7 +4364,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm9
; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u,u,u,u,u,14],zero,ymm1[u,u,u,u,u,15],zero,ymm1[u,u,u,u,u,16],zero,ymm1[u,u,u,u,u,17],zero,ymm1[u,u,u]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
; AVX512-FCP-NEXT: vpor %ymm0, %ymm8, %ymm0
; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = 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]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
@@ -4414,74 +4375,67 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: # ymm10 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm8
; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm8
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm13
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm13[13,u,u,u,u],zero,zero,ymm13[14,u,u,u,u],zero,zero,ymm13[15,u,u,u,u],zero,zero,ymm13[16,u,u,u,u],zero,zero,ymm13[17,u,u]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm12
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm12[13,u,u,u,u],zero,zero,ymm12[14,u,u,u,u],zero,zero,ymm12[15,u,u,u,u],zero,zero,ymm12[16,u,u,u,u],zero,zero,ymm12[17,u,u]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpand %ymm0, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero,ymm3[25]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm4[23,u,u,u],zero,ymm4[26],zero,ymm4[24,u,u,u],zero,ymm4[27],zero
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm0, %zmm9, %zmm0
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm3[19],zero,ymm3[21,20,21,22],zero,ymm3[20],zero,ymm3[22,23]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm10, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vporq %zmm9, %zmm10, %zmm9
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm2[19],zero,ymm2[21,20,21,22],zero,ymm2[20],zero,ymm2[22,23]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm10, %zmm11, %zmm10
+; AVX512-FCP-NEXT: vporq %zmm0, %zmm10, %zmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm10 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,5,4,0,5,0,4,0]
+; AVX512-FCP-NEXT: vpermd %ymm10, %ymm11, %ymm10
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[20],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm0, %zmm11, %zmm0
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm11 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,5,4,0,5,0,4,0]
-; AVX512-FCP-NEXT: vpermd %ymm11, %ymm12, %ymm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm11
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm10
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm6[30],zero,ymm6[28,u,u,u],zero,ymm6[31],zero,ymm6[29,u]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm5, %ymm0
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm5
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512-FCP-NEXT: vpor %ymm4, %ymm3, %ymm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm2[28],zero,ymm2[30,31,30,31],zero,ymm2[29],zero,ymm2[31,28,29]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm9, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm2
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
-; AVX512-FCP-NEXT: vmovdqa %ymm2, 192(%rax)
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa %ymm1, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
@@ -4490,153 +4444,138 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ: # %bb.0:
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm5
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm6
-; AVX512DQ-NEXT: vmovdqa (%r10), %ymm4
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm9 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm8, %zmm9
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm5
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm1
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm2
+; AVX512DQ-NEXT: vmovdqa (%r10), %ymm0
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero,ymm3[25]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm5[18],zero,ymm5[20,21,20,21],zero,ymm5[19],zero,ymm5[19,20,21,22],zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm5[23],zero,ymm5[23,24,25,26],zero,ymm5[24],zero,ymm5[30,31]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm4[23,u,u,u],zero,ymm4[26],zero,ymm4[24,u,u,u],zero,ymm4[27],zero
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
; AVX512DQ-NEXT: vporq %zmm7, %zmm8, %zmm7
-; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512DQ-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-NEXT: vpandq %ymm16, %ymm8, %ymm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm11[18,19,20,21],zero,ymm11[19],zero,ymm11[25,26,27,22],zero,ymm11[20],zero
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm3[19],zero,ymm3[21,20,21,22],zero,ymm3[20],zero,ymm3[22,23]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
; AVX512DQ-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm9, %zmm8, %zmm9
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512DQ-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-NEXT: vpandq %ymm17, %ymm8, %ymm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm11[23],zero,ymm11[21,22,23,26],zero,ymm11[24],zero,ymm11[28,29,26,27]
-; AVX512DQ-NEXT: vmovdqa64 %ymm11, %ymm20
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm10, %zmm8, %zmm8
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm11[u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6,u,u,u],zero
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm12
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm12[u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u,u,9]
-; AVX512DQ-NEXT: vpor %xmm7, %xmm9, %xmm7
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm7, %zmm9, %zmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm10 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm7
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[u,u,u],zero,xmm7[7],zero,xmm7[5,u,u,u],zero,xmm7[8],zero,xmm7[6,u,u]
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm14 = xmm9[u,u,u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u]
-; AVX512DQ-NEXT: vpor %xmm13, %xmm14, %xmm13
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm13, %zmm14, %zmm13
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm18 = zmm13[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm18
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm13
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = zero,xmm13[4,u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6]
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm14
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm15 = xmm14[4],zero,xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero
-; AVX512DQ-NEXT: vpor %xmm10, %xmm15, %xmm10
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm19 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vmovdqa (%r10), %xmm15
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = xmm15[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm15[1,1,0,0,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,0,1,0,4,4,5,4]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm10
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm10
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14,u,u],zero,zero,zero,zero,ymm1[15,u,u],zero,zero,zero,zero,ymm1[16,u,u],zero,zero,zero,zero,ymm1[17,u,u],zero,zero,zero,zero,ymm1[18]
+; AVX512DQ-NEXT: vporq %zmm8, %zmm9, %zmm8
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm10 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm9, %zmm10
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm7
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm7
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm9[u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u,u],zero
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm10
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm10[u,7],zero,xmm10[5],zero,xmm10[u,u,u,8],zero,xmm10[6],zero,xmm10[u,u,u,9]
+; AVX512DQ-NEXT: vpor %xmm8, %xmm11, %xmm8
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm8, %zmm11, %zmm8
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm14[u,u,u],zero,xmm14[7],zero,xmm14[5,u,u,u],zero,xmm14[8],zero,xmm14[6,u,u]
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm12 = xmm15[u,u,u,7],zero,xmm15[5],zero,xmm15[u,u,u,8],zero,xmm15[6],zero,xmm15[u,u]
+; AVX512DQ-NEXT: vpor %xmm11, %xmm12, %xmm11
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm11, %zmm12, %zmm11
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm16 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm16
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm11
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = zero,xmm11[4,u,u,u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6]
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm12
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm13 = xmm12[4],zero,xmm12[u,u,u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero
+; AVX512DQ-NEXT: vpor %xmm8, %xmm13, %xmm8
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm8, %zmm13, %zmm8
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm17 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vmovdqa (%r10), %xmm13
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm13[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm13[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm8
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm8
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm4[14,u,u],zero,zero,zero,zero,ymm4[15,u,u],zero,zero,zero,zero,ymm4[16,u,u],zero,zero,zero,zero,ymm4[17,u,u],zero,zero,zero,zero,ymm4[18]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,14],zero,ymm3[u,u,0,1,14,15],zero,ymm3[u,u,13,2,3,16],zero,ymm3[u,u,28,29,16,17],zero,ymm3[u,u,19,28,29,18],zero
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm15[8],xmm14[9],xmm15[9],xmm14[10],xmm15[10],xmm14[11],xmm15[11],xmm14[12],xmm15[12],xmm14[13],xmm15[13],xmm14[14],xmm15[14],xmm14[15],xmm15[15]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u,u,u]
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm2
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,14],zero,ymm2[u,u,u,u,u,15],zero,ymm2[u,u,u,u,u,16],zero,ymm2[u,u,u,u,u,17],zero,ymm2[u,u,u,u,u]
-; AVX512DQ-NEXT: vpor %ymm1, %ymm7, %ymm1
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm14 = ymm5[u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u,u,u]
+; AVX512DQ-NEXT: vpor %ymm1, %ymm14, %ymm1
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm5[u,u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[13,u,u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u]
-; AVX512DQ-NEXT: vpor %ymm0, %ymm7, %ymm0
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm15[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm7
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm4[13,u,u,u,u],zero,zero,ymm4[14,u,u,u,u],zero,zero,ymm4[15,u,u,u,u],zero,zero,ymm4[16,u,u,u,u],zero,zero,ymm4[17,u,u]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm14
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u,u,u,u,u,14],zero,ymm14[u,u,u,u,u,15],zero,ymm14[u,u,u,u,u,16],zero,ymm14[u,u,u,u,u,17],zero,ymm14[u,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
+; AVX512DQ-NEXT: vpor %ymm0, %ymm9, %ymm0
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm13[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm11
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm11[13,u,u,u,u],zero,zero,ymm11[14,u,u,u,u],zero,zero,ymm11[15,u,u,u,u],zero,zero,ymm11[16,u,u,u,u],zero,zero,ymm11[17,u,u]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm9
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm6[30],zero,ymm6[28,u,u,u],zero,ymm6[31],zero,ymm6[29,u]
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512DQ-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm16, %ymm0, %ymm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512DQ-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm17, %ymm1, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm3
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
-; AVX512DQ-NEXT: vmovdqa %ymm3, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm10, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm2
+; AVX512DQ-NEXT: vmovdqa %ymm2, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm9, 64(%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -4644,12 +4583,12 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP: # %bb.0:
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm3
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm5
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm6
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa64 (%r10), %ymm17
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm8
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm8[u],zero,xmm8[7],zero,xmm8[5,u,u,u],zero,xmm8[8],zero,xmm8[6,u,u,u],zero
@@ -4689,7 +4628,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm7
; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm4[14,u,u],zero,zero,zero,zero,ymm4[15,u,u],zero,zero,zero,zero,ymm4[16,u,u],zero,zero,zero,zero,ymm4[17,u,u],zero,zero,zero,zero,ymm4[18]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm3[0,1,14],zero,ymm3[u,u,0,1,14,15],zero,ymm3[u,u,13,2,3,16],zero,ymm3[u,u,28,29,16,17],zero,ymm3[u,u,19,28,29,18],zero
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm15, %ymm0
; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
@@ -4704,7 +4643,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm9
; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u,u,u,u,u,14],zero,ymm1[u,u,u,u,u,15],zero,ymm1[u,u,u,u,u,16],zero,ymm1[u,u,u,u,u,17],zero,ymm1[u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm8, %ymm0
; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = 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-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
@@ -4715,74 +4654,67 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: # ymm10 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm8
; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm13[13,u,u,u,u],zero,zero,ymm13[14,u,u,u,u],zero,zero,ymm13[15,u,u,u,u],zero,zero,ymm13[16,u,u,u,u],zero,zero,ymm13[17,u,u]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm12
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm12[13,u,u,u,u],zero,zero,ymm12[14,u,u,u,u],zero,zero,ymm12[15,u,u,u,u],zero,zero,ymm12[16,u,u,u,u],zero,zero,ymm12[17,u,u]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpand %ymm0, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27],zero,ymm3[25]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm5[18,19,20,21],zero,ymm5[19],zero,ymm5[25,26,27,22],zero,ymm5[20],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm4[23,u,u,u],zero,ymm4[26],zero,ymm4[24,u,u,u],zero,ymm4[27],zero
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm0, %zmm9, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm3[19],zero,ymm3[21,20,21,22],zero,ymm3[20],zero,ymm3[22,23]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm10, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[21],zero,ymm4[19],zero,zero,zero,zero,ymm4[22],zero,ymm4[20],zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vporq %zmm9, %zmm10, %zmm9
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm2[19],zero,ymm2[21,20,21,22],zero,ymm2[20],zero,ymm2[22,23]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm10, %zmm11, %zmm10
+; AVX512DQ-FCP-NEXT: vporq %zmm0, %zmm10, %zmm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm10 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,5,4,0,5,0,4,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm11, %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[20],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm1[18],zero,ymm1[20,21,20,21],zero,ymm1[19],zero,ymm1[19,20,21,22],zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm1[23],zero,ymm1[23,24,25,26],zero,ymm1[24],zero,ymm1[30,31]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm0, %zmm11, %zmm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm11 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,5,4,0,5,0,4,0]
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm12, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm11
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm10
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm6[30],zero,ymm6[28,u,u,u],zero,ymm6[31],zero,ymm6[29,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm5, %ymm0
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm2[28],zero,ymm2[30,31,30,31],zero,ymm2[29],zero,ymm2[31,28,29]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm9, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, 192(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, 192(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
@@ -4841,33 +4773,29 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm0 {%k1}
; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm9
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm9[18,19,20,21],zero,zmm9[19],zero,zmm9[25,26,27,22],zero,zmm9[20],zero,zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero,zmm9[59],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm15
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[18],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero,zmm15[59],zero,zmm15[57]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: vporq %zmm9, %zmm15, %zmm9
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
; AVX512BW-NEXT: movl $676341840, %ecx # imm = 0x28502850
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm15 {%k1} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm16 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm1[23],zero,ymm1[21,22,23,26],zero,ymm1[24],zero,ymm1[28,29,26,27]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,3,2,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm17 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
; AVX512BW-NEXT: vporq %ymm16, %ymm17, %ymm16
; AVX512BW-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
; AVX512BW-NEXT: kmovq %rcx, %k2
; AVX512BW-NEXT: vmovdqu8 %zmm15, %zmm9 {%k2}
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
; AVX512BW-NEXT: vpermw %zmm7, %zmm15, %zmm15
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: vporq %zmm16, %zmm17, %zmm16
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
; AVX512BW-NEXT: kmovq %rcx, %k2
; AVX512BW-NEXT: vmovdqu8 %zmm15, %zmm16 {%k2}
@@ -4924,10 +4852,9 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermw %ymm7, %ymm2, %ymm2
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX512BW-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512BW-NEXT: movl $-2130574328, %ecx # imm = 0x81020408
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu8 %ymm2, %ymm3 {%k1}
@@ -4945,8 +4872,8 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP: # %bb.0:
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm2
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
@@ -4958,8 +4885,8 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[0,1,14],zero,ymm3[12,13,0,1,14,15],zero,ymm3[3,12,13,2,3,16],zero,ymm3[30,31,28,29,16,17],zero,ymm3[31,18,19,28,29,18],zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero,zero,ymm4[18]
; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm6, %ymm0
; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm14
@@ -4993,30 +4920,27 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm0 {%k1}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm10
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm10
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,zmm10[19],zero,zmm10[21,20,21,22],zero,zmm10[20],zero,zmm10[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm15
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm15
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20],zero,zero,zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm15[57],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vporq %zmm10, %zmm15, %zmm15
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm10
+; AVX512BW-FCP-NEXT: vporq %zmm10, %zmm15, %zmm10
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm15 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm10
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm10[55],zero,zero,zero,zero,zmm10[58],zero,zmm10[56],zero,zero,zero,zero,zmm10[59],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm16
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm16
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[18],zero,zero,zero,zero,zmm16[21],zero,zmm16[19],zero,zero,zero,zero,zmm16[22],zero,zmm16[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[55],zero,zero,zero,zero,zmm16[58],zero,zmm16[56],zero,zero,zero,zero,zmm16[59],zero,zmm16[57]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm10, %zmm16, %zmm10
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm10 {%k1}
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
; AVX512BW-FCP-NEXT: vpermw %zmm7, %zmm15, %zmm15
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm16, %zmm17, %zmm16
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm16 {%k1}
@@ -5055,16 +4979,14 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm8 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm3, %ymm3
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512BW-FCP-NEXT: movl $101455920, %ecx # imm = 0x60C1830
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm1 {%k1}
@@ -5072,10 +4994,9 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512BW-FCP-NEXT: vpermw %ymm7, %ymm2, %ymm2
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512BW-FCP-NEXT: movl $-2130574328, %ecx # imm = 0x81020408
; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %ymm2, %ymm3 {%k1}
@@ -5143,33 +5064,29 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm0 {%k1}
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm9
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm9[18,19,20,21],zero,zmm9[19],zero,zmm9[25,26,27,22],zero,zmm9[20],zero,zmm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm9[55],zero,zero,zero,zero,zmm9[58],zero,zmm9[56],zero,zero,zero,zero,zmm9[59],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm15
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[18],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero,zmm15[59],zero,zmm15[57]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: vporq %zmm9, %zmm15, %zmm9
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,0,1,1,4,4,5,5]
; AVX512DQ-BW-NEXT: movl $676341840, %ecx # imm = 0x28502850
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm15 {%k1} = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm16 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm1[23],zero,ymm1[21,22,23,26],zero,ymm1[24],zero,ymm1[28,29,26,27]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,3,2,3]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm17 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
; AVX512DQ-BW-NEXT: vporq %ymm16, %ymm17, %ymm16
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
; AVX512DQ-BW-NEXT: kmovq %rcx, %k2
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm15, %zmm9 {%k2}
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
; AVX512DQ-BW-NEXT: vpermw %zmm7, %zmm15, %zmm15
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: vporq %zmm16, %zmm17, %zmm16
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
; AVX512DQ-BW-NEXT: kmovq %rcx, %k2
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm15, %zmm16 {%k2}
@@ -5226,10 +5143,9 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512DQ-BW-NEXT: vpermw %ymm7, %ymm2, %ymm2
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX512DQ-BW-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512DQ-BW-NEXT: movl $-2130574328, %ecx # imm = 0x81020408
; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm2, %ymm3 {%k1}
@@ -5247,8 +5163,8 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP: # %bb.0:
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm2
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[0,1,0,1,14],zero,ymm1[14,15,0,1,14,15],zero,ymm1[13,14,15,16,17,16],zero,ymm1[30,31,30,31,16,17],zero,ymm1[31,28,29,30,31]
@@ -5260,8 +5176,8 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm5
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[0,1,14],zero,ymm3[12,13,0,1,14,15],zero,ymm3[3,12,13,2,3,16],zero,ymm3[30,31,28,29,16,17],zero,ymm3[31,18,19,28,29,18],zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,ymm4[14],zero,zero,zero,zero,zero,zero,ymm4[15],zero,zero,zero,zero,zero,zero,ymm4[16],zero,zero,zero,zero,zero,zero,ymm4[17],zero,zero,zero,zero,zero,zero,ymm4[18]
; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm6, %ymm0
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm14
@@ -5295,30 +5211,27 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm0 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm10
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm10
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,zmm10[19],zero,zmm10[21,20,21,22],zero,zmm10[20],zero,zmm10[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm10[55],zero,zmm10[53,54,55,58],zero,zmm10[56],zero,zmm10[60,61,58,59]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm15
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm15
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm15 = zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm15[21],zero,zmm15[19],zero,zero,zero,zero,zmm15[22],zero,zmm15[20],zero,zero,zmm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm15[57],zero,zmm15[55],zero,zero,zero,zero,zmm15[58],zero,zmm15[56],zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm15 = zmm15[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vporq %zmm10, %zmm15, %zmm15
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm10
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm10, %zmm15, %zmm10
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm15 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm10
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm10 = zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm10[18,19,20,21],zero,zmm10[19],zero,zmm10[25,26,27,22],zero,zmm10[20],zero,zmm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm10[55],zero,zero,zero,zero,zmm10[58],zero,zmm10[56],zero,zero,zero,zero,zmm10[59],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm16
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm16
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[18],zero,zero,zero,zero,zmm16[21],zero,zmm16[19],zero,zero,zero,zero,zmm16[22],zero,zmm16[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[55],zero,zero,zero,zero,zmm16[58],zero,zmm16[56],zero,zero,zero,zero,zmm16[59],zero,zmm16[57]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm10, %zmm16, %zmm10
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $-9005497107459067808, %rcx # imm = 0x83060C180C183060
; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm10 {%k1}
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm15 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
; AVX512DQ-BW-FCP-NEXT: vpermw %zmm7, %zmm15, %zmm15
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm16 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm6[18],zero,zmm6[20,21,20,21],zero,zmm6[19],zero,zmm6[19,20,21,22],zero,zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm6[55],zero,zmm6[55,56,57,58],zero,zmm6[56],zero,zmm6[62,63]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm17 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[20],zero,zmm5[18],zero,zero,zero,zero,zmm5[21],zero,zmm5[19],zero,zero,zero,zero,zmm5[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[57],zero,zmm5[55],zero,zero,zero,zero,zmm5[58],zero,zmm5[56],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm16, %zmm17, %zmm16
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $1161999626690365456, %rcx # imm = 0x1020408102040810
; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm16 {%k1}
@@ -5357,16 +5270,14 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm9, %zmm8 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm3, %ymm3
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: movl $101455920, %ecx # imm = 0x60C1830
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm3, %ymm1 {%k1}
@@ -5374,10 +5285,9 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512DQ-BW-FCP-NEXT: vpermw %ymm7, %ymm2, %ymm2
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
; AVX512DQ-BW-FCP-NEXT: movl $-2130574328, %ecx # imm = 0x81020408
; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %ymm2, %ymm3 {%k1}
@@ -7376,8 +7286,8 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT: vmovdqa 32(%rdi), %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa 32(%rcx), %ymm7
@@ -7389,15 +7299,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqa 32(%rax), %ymm3
; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,ymm0[27,20,21,26],zero,ymm0[24],zero,ymm0[26,27,26,27],zero,ymm0[25]
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero,ymm1[27],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero
; AVX2-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[25],zero,ymm7[23],zero,zero,zero,zero,ymm7[26],zero,ymm7[24],zero,zero,zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27]
-; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,24,25,26,27,24,25,30,31]
@@ -7415,17 +7323,16 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vmovdqa (%r8), %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,ymm0[27,28,29,30],zero,ymm0[28],zero,ymm0[26,27,30,31],zero,ymm0[29]
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-NEXT: vmovdqa (%r9), %ymm1
; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,ymm0[27,28,29,30],zero,ymm0[28],zero,ymm0[26,27,30,31],zero,ymm0[29]
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vmovdqa (%rax), %ymm1
; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm2 = [u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0]
; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-NEXT: vmovdqa (%rdx), %ymm1
@@ -7622,12 +7529,11 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT: vpshufb %ymm3, %ymm0, %ymm2
-; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-NEXT: vpshufb %ymm4, %ymm0, %ymm5
-; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vpor %ymm2, %ymm5, %ymm2
+; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-NEXT: vpshuflw $150, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
; AVX2-NEXT: # ymm5 = mem[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX2-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[0,1,1,3,4,5,5,7]
@@ -7635,15 +7541,14 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
; AVX2-NEXT: vpblendvb %ymm6, %ymm2, %ymm5, %ymm0
; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX2-NEXT: vpshufb %ymm3, %ymm14, %ymm3
-; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX2-NEXT: vpshufb %ymm3, %ymm8, %ymm3
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
; AVX2-NEXT: vpshufb %ymm4, %ymm13, %ymm4
-; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX2-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX2-NEXT: vpshuflw {{.*#+}} ymm4 = ymm8[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX2-NEXT: vpshuflw {{.*#+}} ymm4 = ymm14[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,1,1,3,4,5,5,7]
; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,2]
; AVX2-NEXT: vpblendvb %ymm6, %ymm3, %ymm4, %ymm0
@@ -7665,24 +7570,22 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vpblendvb %ymm7, %ymm6, %ymm5, %ymm5
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
; AVX2-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX2-NEXT: vpshufb %ymm6, %ymm10, %ymm7
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-NEXT: vpshufb %ymm3, %ymm2, %ymm9
-; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX2-NEXT: vpshufb %ymm6, %ymm2, %ymm7
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
+; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX2-NEXT: vpshufb %ymm3, %ymm10, %ymm9
; AVX2-NEXT: vpor %ymm7, %ymm9, %ymm7
+; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
; AVX2-NEXT: vpblendvb %ymm9, %ymm7, %ymm4, %ymm4
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-NEXT: vpshufb %ymm6, %ymm15, %ymm6
-; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-NEXT: vpshufb %ymm3, %ymm12, %ymm7
-; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vpblendvb %ymm9, %ymm6, %ymm5, %ymm5
; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
@@ -7693,26 +7596,22 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vpblendvb %ymm6, %ymm5, %ymm3, %ymm3
; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm0[23],zero,ymm0[27,20,21,26],zero,ymm0[24],zero,ymm0[26,27,26,27],zero,ymm0[25]
-; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
; AVX2-NEXT: vmovdqa %ymm1, %ymm0
; AVX2-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero,ymm1[27],zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX2-NEXT: vpshufb {{.*#+}} ymm5 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm15[25],zero,ymm15[23],zero,zero,zero,zero,ymm15[26],zero,ymm15[24],zero,zero,zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[25],zero,ymm12[23],zero,zero,zero,zero,ymm12[26],zero,ymm12[24],zero,zero,zero,zero,ymm12[27]
-; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
; AVX2-NEXT: vpor %ymm5, %ymm6, %ymm5
+; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-NEXT: vpblendvb %ymm3, %ymm4, %ymm5, %ymm4
-; AVX2-NEXT: vpshufb {{.*#+}} ymm5 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[25],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX2-NEXT: vpshufb {{.*#+}} ymm5 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[25],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero
; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm13[25],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero
-; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
; AVX2-NEXT: vpor %ymm5, %ymm6, %ymm5
-; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX2-NEXT: vmovdqa %ymm8, %ymm14
+; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
+; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
; AVX2-NEXT: vpblendvb %ymm7, %ymm5, %ymm6, %ymm5
; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
@@ -7725,10 +7624,11 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vpshufb %ymm7, %ymm4, %ymm8
; AVX2-NEXT: vpor %ymm5, %ymm8, %ymm5
; AVX2-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX2-NEXT: vpshufb %ymm8, %ymm10, %ymm9
-; AVX2-NEXT: vmovdqa %ymm10, %ymm3
-; AVX2-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,0,128,14,128,128,128,128,1,128,15,128,128,128,128,128,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128]
-; AVX2-NEXT: vpshufb %ymm10, %ymm2, %ymm11
+; AVX2-NEXT: vpshufb %ymm8, %ymm2, %ymm9
+; AVX2-NEXT: vmovdqa %ymm2, %ymm3
+; AVX2-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,0,128,14,128,128,128,128,1,128,15,128,128,128,128,128,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128]
+; AVX2-NEXT: vpshufb %ymm1, %ymm10, %ymm11
+; AVX2-NEXT: vmovdqa %ymm10, %ymm2
; AVX2-NEXT: vpor %ymm9, %ymm11, %ymm9
; AVX2-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255]
; AVX2-NEXT: vpblendvb %ymm11, %ymm5, %ymm9, %ymm5
@@ -7737,7 +7637,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-NEXT: vpshufb %ymm7, %ymm0, %ymm7
; AVX2-NEXT: vpor %ymm6, %ymm7, %ymm6
; AVX2-NEXT: vpshufb %ymm8, %ymm15, %ymm7
-; AVX2-NEXT: vpshufb %ymm10, %ymm12, %ymm8
+; AVX2-NEXT: vpshufb %ymm1, %ymm12, %ymm8
; AVX2-NEXT: vpor %ymm7, %ymm8, %ymm7
; AVX2-NEXT: vpblendvb %ymm11, %ymm6, %ymm7, %ymm6
; AVX2-NEXT: vmovdqa {{.*#+}} ymm7 = [128,1,2,3,0,128,14,128,0,1,0,1,128,15,128,15,128,17,18,19,16,128,30,128,16,17,16,17,128,31,128,31]
@@ -7820,7 +7720,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX2-FP-LABEL: store_i8_stride7_vf64:
; AVX2-FP: # %bb.0:
-; AVX2-FP-NEXT: subq $648, %rsp # imm = 0x288
+; AVX2-FP-NEXT: subq $616, %rsp # imm = 0x268
; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm1
; AVX2-FP-NEXT: vmovdqa 32(%rsi), %ymm7
@@ -7832,20 +7732,18 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[23],zero,ymm1[27,20,21,26],zero,ymm1[24],zero,ymm1[26,27,26,27],zero,ymm1[25]
; AVX2-FP-NEXT: vmovdqa %ymm1, %ymm8
; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm7[23],zero,zero,zero,zero,ymm7[26],zero,ymm7[24],zero,zero,zero,zero,ymm7[27],zero
; AVX2-FP-NEXT: vmovdqa %ymm7, %ymm9
; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm7
; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27]
; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,24,25,26,27,24,25,30,31]
@@ -7865,15 +7763,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm6[17,18,19,30],zero,ymm6[28],zero,ymm6[28,29,30,31],zero,ymm6[29],zero,ymm6[31]
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-FP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm2 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,26,27,30,31,30,31,28,29,28,29,28,29,28,29]
@@ -8052,184 +7948,167 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vmovdqa (%rdx), %ymm1
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[17,18,19,30],zero,ymm1[28],zero,ymm1[28,29,30,31],zero,ymm1[29],zero,ymm1[31]
-; AVX2-FP-NEXT: vmovdqa %ymm1, %ymm4
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm10
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX2-FP-NEXT: vmovdqa %ymm1, %ymm3
+; AVX2-FP-NEXT: vmovdqa (%rcx), %ymm5
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero
; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm2
-; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm1
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX2-FP-NEXT: vmovdqa %ymm1, %ymm12
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm5
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm3, %ymm6, %ymm3
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
-; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm6
-; AVX2-FP-NEXT: vmovdqa (%r8), %ymm0
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,ymm0[27,28,29,30],zero,ymm0[28],zero,ymm0[26,27,30,31],zero,ymm0[29]
-; AVX2-FP-NEXT: vmovdqa %ymm0, %ymm1
-; AVX2-FP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm0
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm4 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero,ymm0[29],zero,zero,zero
+; AVX2-FP-NEXT: vmovdqa %ymm0, %ymm11
+; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm13
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29],zero,zero,zero,zero
+; AVX2-FP-NEXT: vpor %ymm4, %ymm6, %ymm4
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa (%r9), %ymm0
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27],zero,zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero
-; AVX2-FP-NEXT: vmovdqa %ymm0, %ymm3
-; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
+; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm2, %ymm4, %ymm6
+; AVX2-FP-NEXT: vmovdqa (%r8), %ymm14
+; AVX2-FP-NEXT: vmovdqa (%r9), %ymm1
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,ymm14[27,28,29,30],zero,ymm14[28],zero,ymm14[26,27,30,31],zero,ymm14[29]
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero
+; AVX2-FP-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
; AVX2-FP-NEXT: vpor %ymm2, %ymm7, %ymm7
; AVX2-FP-NEXT: vmovdqa (%rax), %ymm0
; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX2-FP-NEXT: vmovdqa %ymm0, %ymm11
+; AVX2-FP-NEXT: vmovdqa %ymm0, %ymm10
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0]
; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0]
; AVX2-FP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm0
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[23],zero,ymm5[27,20,21,26],zero,ymm5[24],zero,ymm5[26,27,26,27],zero,ymm5[25]
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[23],zero,ymm13[27,20,21,26],zero,ymm13[24],zero,ymm13[26,27,26,27],zero,ymm13[25]
+; AVX2-FP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero,zero,zero,ymm11[27],zero
+; AVX2-FP-NEXT: vmovdqa %ymm11, %ymm2
+; AVX2-FP-NEXT: vpor %ymm6, %ymm8, %ymm6
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm5[25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero
+; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27]
+; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm12[23],zero,zero,zero,zero,ymm12[26],zero,ymm12[24],zero,zero,zero,zero,ymm12[27],zero
-; AVX2-FP-NEXT: vmovdqa %ymm12, %ymm14
-; AVX2-FP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm10[25],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero,zero,zero
-; AVX2-FP-NEXT: vmovdqa %ymm10, %ymm13
-; AVX2-FP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa %ymm4, %ymm2
-; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero,ymm4[27]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm7, %ymm8, %ymm7
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
-; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm6, %ymm7, %ymm6
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm6, %ymm8, %ymm6
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm14[25],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero,zero
+; AVX2-FP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm9 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
-; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
-; AVX2-FP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm0
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
+; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm8
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
+; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm6, %ymm8, %ymm0
; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
-; AVX2-FP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX2-FP-NEXT: # ymm9 = mem[0,1,0,1]
; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm0, %ymm7
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm6, %ymm10
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm7, %ymm10, %ymm7
-; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX2-FP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm4, %ymm11
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128]
+; AVX2-FP-NEXT: vpshufb %ymm9, %ymm0, %ymm8
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm0, %ymm11
+; AVX2-FP-NEXT: vpor %ymm8, %ymm11, %ymm8
+; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX2-FP-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm11, %ymm7, %ymm12
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128]
; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm12, %ymm1, %ymm15
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm11, %ymm15, %ymm11
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
-; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm7, %ymm11, %ymm0
-; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpshufb %ymm8, %ymm13, %ymm8
+; AVX2-FP-NEXT: vpshufb %ymm0, %ymm1, %ymm15
+; AVX2-FP-NEXT: vpor %ymm12, %ymm15, %ymm12
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm14, %ymm9
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
+; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm8, %ymm12, %ymm4
+; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpshufb %ymm9, %ymm5, %ymm9
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm3, %ymm10
+; AVX2-FP-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm3
+; AVX2-FP-NEXT: vpshufb %ymm11, %ymm2, %ymm10
+; AVX2-FP-NEXT: vpshufb %ymm0, %ymm13, %ymm11
+; AVX2-FP-NEXT: vpor %ymm10, %ymm11, %ymm10
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX2-FP-NEXT: vpshufb %ymm12, %ymm5, %ymm10
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm9, %ymm10, %ymm9
-; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm8, %ymm9, %ymm8
-; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX2-FP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm14, %ymm10
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm11, %ymm3, %ymm12
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm10, %ymm12, %ymm10
-; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [20,21,18,19,18,19,20,21,18,19,20,21,28,29,30,31,20,21,18,19,18,19,20,21,18,19,20,21,28,29,30,31]
-; AVX2-FP-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm12, %ymm13, %ymm15
+; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm9, %ymm10, %ymm9
+; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX2-FP-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm8, %ymm11
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm12, %ymm4, %ymm13
+; AVX2-FP-NEXT: vpor %ymm11, %ymm13, %ymm11
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [20,21,18,19,18,19,20,21,18,19,20,21,28,29,30,31,20,21,18,19,18,19,20,21,18,19,20,21,28,29,30,31]
+; AVX2-FP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm13, %ymm6, %ymm15
; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,2]
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
-; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm10, %ymm15, %ymm10
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm9, %ymm15, %ymm9
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX2-FP-NEXT: vmovdqu (%rsp), %ymm2 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm11, %ymm2, %ymm11
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX2-FP-NEXT: vpor %ymm9, %ymm11, %ymm9
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm12, %ymm7, %ymm11
-; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,2]
-; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm9, %ymm11, %ymm0
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
-; AVX2-FP-NEXT: vpblendvb %ymm9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
-; AVX2-FP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FP-NEXT: vpblendvb %ymm9, %ymm8, %ymm0, %ymm8
+; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm11, %ymm15, %ymm11
+; AVX2-FP-NEXT: vmovdqu (%rsp), %ymm15 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm10, %ymm15, %ymm10
+; AVX2-FP-NEXT: vpshufb %ymm12, %ymm14, %ymm12
+; AVX2-FP-NEXT: vpor %ymm10, %ymm12, %ymm10
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm13, %ymm2, %ymm12
+; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,2]
+; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm10, %ymm12, %ymm0
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
+; AVX2-FP-NEXT: vpblendvb %ymm10, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload
+; AVX2-FP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FP-NEXT: vpblendvb %ymm10, %ymm9, %ymm0, %ymm9
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX2-FP-NEXT: vpshufb %ymm0, %ymm4, %ymm9
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,128,14,128,128,128,128,1,128,15,128,128,128,128,2,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128,18,128]
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm1, %ymm11
-; AVX2-FP-NEXT: vpor %ymm9, %ymm11, %ymm9
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm0, %ymm7, %ymm10
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,128,14,128,128,128,128,1,128,15,128,128,128,128,2,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128,18,128]
; AVX2-FP-NEXT: vpshufb %ymm11, %ymm1, %ymm12
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,0,128,14,128,128,128,128,1,128,15,128,128,128,128,128,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128]
-; AVX2-FP-NEXT: vpshufb %ymm4, %ymm6, %ymm6
-; AVX2-FP-NEXT: vpor %ymm6, %ymm12, %ymm6
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255]
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm9, %ymm6, %ymm9
+; AVX2-FP-NEXT: vpor %ymm10, %ymm12, %ymm10
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
-; AVX2-FP-NEXT: vpshufb %ymm10, %ymm5, %ymm1
-; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FP-NEXT: vpshufb %ymm12, %ymm1, %ymm13
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,0,128,14,128,128,128,128,1,128,15,128,128,128,128,128,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128]
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm14, %ymm1, %ymm7
+; AVX2-FP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255]
+; AVX2-FP-NEXT: vpblendvb %ymm13, %ymm10, %ymm7, %ymm10
+; AVX2-FP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
-; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX2-FP-NEXT: vpshufb %ymm4, %ymm5, %ymm4
-; AVX2-FP-NEXT: vpor %ymm1, %ymm4, %ymm1
-; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm0, %ymm1, %ymm0
+; AVX2-FP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FP-NEXT: vpshufb %ymm12, %ymm5, %ymm1
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm14, %ymm3, %ymm3
+; AVX2-FP-NEXT: vpor %ymm1, %ymm3, %ymm1
+; AVX2-FP-NEXT: vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,1,2,3,0,128,14,128,0,1,0,1,128,15,128,15,128,17,18,19,16,128,30,128,16,17,16,17,128,31,128,31]
-; AVX2-FP-NEXT: vpshufb %ymm1, %ymm3, %ymm4
+; AVX2-FP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm5 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX2-FP-NEXT: vpshufb %ymm5, %ymm14, %ymm6
-; AVX2-FP-NEXT: vpor %ymm4, %ymm6, %ymm4
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX2-FP-NEXT: vpshufb %ymm6, %ymm13, %ymm10
-; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u]
-; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm4, %ymm10, %ymm4
-; AVX2-FP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX2-FP-NEXT: vpshufb %ymm5, %ymm15, %ymm3
-; AVX2-FP-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX2-FP-NEXT: vpshufb %ymm6, %ymm7, %ymm2
-; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm1, %ymm2, %ymm1
+; AVX2-FP-NEXT: vpshufb %ymm5, %ymm8, %ymm7
+; AVX2-FP-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm7 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX2-FP-NEXT: vpshufb %ymm7, %ymm6, %ymm11
+; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u]
+; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm3, %ymm11, %ymm3
+; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX2-FP-NEXT: vpshufb %ymm1, %ymm4, %ymm1
+; AVX2-FP-NEXT: vpshufb %ymm5, %ymm15, %ymm4
+; AVX2-FP-NEXT: vpor %ymm1, %ymm4, %ymm1
+; AVX2-FP-NEXT: vpshufb %ymm7, %ymm2, %ymm2
+; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm1, %ymm2, %ymm1
; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255]
-; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm9, %ymm4, %ymm3
+; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm10, %ymm3, %ymm3
; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT: vmovdqa %ymm0, 96(%rax)
; AVX2-FP-NEXT: vmovdqa %ymm3, 320(%rax)
-; AVX2-FP-NEXT: vmovdqa %ymm8, 128(%rax)
+; AVX2-FP-NEXT: vmovdqa %ymm9, 128(%rax)
; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rax)
; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
@@ -8252,13 +8131,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FP-NEXT: vmovaps %ymm0, 416(%rax)
; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FP-NEXT: vmovaps %ymm0, 384(%rax)
-; AVX2-FP-NEXT: addq $648, %rsp # imm = 0x288
+; AVX2-FP-NEXT: addq $616, %rsp # imm = 0x268
; AVX2-FP-NEXT: vzeroupper
; AVX2-FP-NEXT: retq
;
; AVX2-FCP-LABEL: store_i8_stride7_vf64:
; AVX2-FCP: # %bb.0:
-; AVX2-FCP-NEXT: subq $648, %rsp # imm = 0x288
+; AVX2-FCP-NEXT: subq $616, %rsp # imm = 0x268
; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %ymm7
@@ -8270,20 +8149,18 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[23],zero,ymm1[27,20,21,26],zero,ymm1[24],zero,ymm1[26,27,26,27],zero,ymm1[25]
; AVX2-FCP-NEXT: vmovdqa %ymm1, %ymm8
; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm7[23],zero,zero,zero,zero,ymm7[26],zero,ymm7[24],zero,zero,zero,zero,ymm7[27],zero
; AVX2-FCP-NEXT: vmovdqa %ymm7, %ymm9
; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
; AVX2-FCP-NEXT: vmovdqa %ymm2, %ymm7
; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27]
; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,24,25,26,27,24,25,30,31]
@@ -8303,15 +8180,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm6[17,18,19,30],zero,ymm6[28],zero,ymm6[28,29,30,31],zero,ymm6[29],zero,ymm6[31]
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
; AVX2-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,26,27,30,31,30,31,28,29,28,29,28,29,28,29]
@@ -8332,7 +8207,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX2-FCP-NEXT: vmovdqa %xmm1, %xmm14
-; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
@@ -8349,7 +8224,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm12
; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
; AVX2-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX2-FCP-NEXT: vmovdqa %xmm11, (%rsp) # 16-byte Spill
+; AVX2-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FCP-NEXT: vmovdqa (%rcx), %xmm2
; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -8445,7 +8320,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm1, %ymm0, %ymm0
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm7[8],xmm10[9],xmm7[9],xmm10[10],xmm7[10],xmm10[11],xmm7[11],xmm10[12],xmm7[12],xmm10[13],xmm7[13],xmm10[14],xmm7[14],xmm10[15],xmm7[15]
-; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-FCP-NEXT: vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm1 = xmm1[8],mem[8],xmm1[9],mem[9],xmm1[10],mem[10],xmm1[11],mem[11],xmm1[12],mem[12],xmm1[13],mem[13],xmm1[14],mem[14],xmm1[15],mem[15]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
@@ -8458,7 +8333,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX2-FCP-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload
+; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FCP-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
@@ -8488,186 +8363,167 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vmovdqa (%rdx), %ymm1
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[17,18,19,30],zero,ymm1[28],zero,ymm1[28,29,30,31],zero,ymm1[29],zero,ymm1[31]
-; AVX2-FCP-NEXT: vmovdqa %ymm1, %ymm4
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero
-; AVX2-FCP-NEXT: vmovdqa %ymm2, %ymm10
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX2-FCP-NEXT: vmovdqa %ymm1, %ymm3
+; AVX2-FCP-NEXT: vmovdqa (%rcx), %ymm5
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero
; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm2
-; AVX2-FCP-NEXT: vmovdqa (%rsi), %ymm5
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero,ymm0[29],zero,zero,zero,zero
-; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm12
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
-; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm6
-; AVX2-FCP-NEXT: vmovdqa (%r8), %ymm0
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[27],zero,ymm0[27,28,29,30],zero,ymm0[28],zero,ymm0[26,27,30,31],zero,ymm0[29]
-; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm1
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa (%rsi), %ymm0
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero,ymm0[29],zero,zero,zero
+; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm11
+; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm13
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29],zero,zero,zero,zero
+; AVX2-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27],zero,zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero
-; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm3
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u]
+; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm2, %ymm4, %ymm6
+; AVX2-FCP-NEXT: vmovdqa (%r8), %ymm14
+; AVX2-FCP-NEXT: vmovdqa (%r9), %ymm1
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,ymm14[27,28,29,30],zero,ymm14[28],zero,ymm14[26,27,30,31],zero,ymm14[29]
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero
+; AVX2-FCP-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
; AVX2-FCP-NEXT: vpor %ymm2, %ymm7, %ymm7
; AVX2-FCP-NEXT: vmovdqa (%rax), %ymm0
; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm11
-; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX2-FCP-NEXT: vmovdqa %ymm0, %ymm10
+; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0]
; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0]
; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm0
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[23],zero,ymm12[27,20,21,26],zero,ymm12[24],zero,ymm12[26,27,26,27],zero,ymm12[25]
-; AVX2-FCP-NEXT: vmovdqa %ymm12, %ymm13
-; AVX2-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[23],zero,ymm13[27,20,21,26],zero,ymm13[24],zero,ymm13[26,27,26,27],zero,ymm13[25]
+; AVX2-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero,zero,zero,ymm11[27],zero
+; AVX2-FCP-NEXT: vmovdqa %ymm11, %ymm2
+; AVX2-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm5[25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero
+; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero,ymm3[27]
+; AVX2-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero,ymm5[27],zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm10[25],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero,zero,zero
-; AVX2-FCP-NEXT: vmovdqa %ymm10, %ymm14
-; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa %ymm4, %ymm2
-; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero,ymm4[27]
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0]
-; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm6, %ymm7, %ymm6
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm6, %ymm8, %ymm6
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm14[25],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero,zero
+; AVX2-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
-; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm7, %ymm8, %ymm7
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm8, %ymm6, %ymm7, %ymm0
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u]
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm8, %ymm9, %ymm8
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255]
+; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm6, %ymm8, %ymm0
; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
-; AVX2-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX2-FCP-NEXT: # ymm9 = mem[0,1,0,1]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm7
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm10
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm7, %ymm10, %ymm7
-; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX2-FCP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm11
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128]
+; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm8
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm11
+; AVX2-FCP-NEXT: vpor %ymm8, %ymm11, %ymm8
+; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX2-FCP-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm12
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm15
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm11, %ymm15, %ymm11
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
-; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm7, %ymm11, %ymm0
-; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm14, %ymm8
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm15
+; AVX2-FCP-NEXT: vpor %ymm12, %ymm15, %ymm12
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm9
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm10
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm9, %ymm10, %ymm9
-; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm8, %ymm9, %ymm8
-; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX2-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm14, %ymm10
-; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm12
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm10, %ymm12, %ymm10
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} ymm12 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [4,5,4,5,5,7,4,5]
-; AVX2-FCP-NEXT: vpermd %ymm12, %ymm15, %ymm12
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm10, %ymm12, %ymm10
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm9
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u,u,0,0,255,255,u,u]
+; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm8, %ymm12, %ymm4
+; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm9
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm10
+; AVX2-FCP-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX2-FCP-NEXT: vmovdqa %ymm2, %ymm3
+; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm11
+; AVX2-FCP-NEXT: vpor %ymm10, %ymm11, %ymm10
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm9, %ymm10, %ymm9
+; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX2-FCP-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm11
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128]
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm11
+; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm13
+; AVX2-FCP-NEXT: vpor %ymm11, %ymm13, %ymm11
; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX2-FCP-NEXT: vpor %ymm9, %ymm11, %ymm9
-; AVX2-FCP-NEXT: vmovdqu (%rsp), %ymm7 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshuflw {{.*#+}} ymm11 = ymm7[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX2-FCP-NEXT: vpermd %ymm11, %ymm15, %ymm11
-; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm9, %ymm11, %ymm0
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
-; AVX2-FCP-NEXT: vpblendvb %ymm9, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
-; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-FCP-NEXT: vpblendvb %ymm9, %ymm8, %ymm0, %ymm8
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshuflw {{.*#+}} ymm13 = ymm6[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [4,5,4,5,5,7,4,5]
+; AVX2-FCP-NEXT: vpermd %ymm13, %ymm15, %ymm13
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255]
+; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm11, %ymm13, %ymm11
+; AVX2-FCP-NEXT: vmovdqu (%rsp), %ymm4 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm14, %ymm12
+; AVX2-FCP-NEXT: vpor %ymm10, %ymm12, %ymm10
+; AVX2-FCP-NEXT: vpshuflw $150, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: # ymm12 = mem[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX2-FCP-NEXT: vpermd %ymm12, %ymm15, %ymm12
+; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm10, %ymm12, %ymm0
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0]
+; AVX2-FCP-NEXT: vpblendvb %ymm10, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm14 # 32-byte Folded Reload
+; AVX2-FCP-NEXT: vpblendvb %ymm10, %ymm9, %ymm0, %ymm9
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm9
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,128,14,128,128,128,128,1,128,15,128,128,128,128,2,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128,18,128]
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm11
-; AVX2-FCP-NEXT: vpor %ymm9, %ymm11, %ymm9
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm10
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,128,14,128,128,128,128,1,128,15,128,128,128,128,2,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128,18,128]
; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm12
+; AVX2-FCP-NEXT: vpor %ymm10, %ymm12, %ymm10
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm13
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,0,128,14,128,128,128,128,1,128,15,128,128,128,128,128,128,16,128,30,128,128,128,128,17,128,31,128,128,128,128]
-; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm6
-; AVX2-FCP-NEXT: vpor %ymm6, %ymm12, %ymm6
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm9, %ymm6, %ymm9
-; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm7
+; AVX2-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255,0,0,u,u,u,255,255]
+; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm10, %ymm7, %ymm10
+; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
-; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm4, %ymm4
-; AVX2-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
-; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm0, %ymm1, %ymm0
+; AVX2-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm1
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm3
+; AVX2-FCP-NEXT: vpor %ymm1, %ymm3, %ymm1
+; AVX2-FCP-NEXT: vpblendvb %ymm13, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,1,2,3,0,128,14,128,0,1,0,1,128,15,128,15,128,17,18,19,16,128,30,128,16,17,16,17,128,31,128,31]
-; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm4
+; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm3
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm14, %ymm6
-; AVX2-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm10
-; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u]
-; AVX2-FCP-NEXT: vpblendvb %ymm11, %ymm4, %ymm10, %ymm4
+; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm8, %ymm7
+; AVX2-FCP-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm6, %ymm11
+; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u,u,u,255,255,0,u,u]
+; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm3, %ymm11, %ymm3
+; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm4, %ymm4
+; AVX2-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm3
-; AVX2-FCP-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm2
-; AVX2-FCP-NEXT: vpblendvb %ymm11, %ymm1, %ymm2, %ymm1
+; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2
+; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm1, %ymm2, %ymm1
; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255,255,255,0,0,0,255,255]
-; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm9, %ymm4, %ymm3
+; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm10, %ymm3, %ymm3
; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT: vmovdqa %ymm0, 96(%rax)
; AVX2-FCP-NEXT: vmovdqa %ymm3, 320(%rax)
-; AVX2-FCP-NEXT: vmovdqa %ymm8, 128(%rax)
-; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rax)
+; AVX2-FCP-NEXT: vmovdqa %ymm9, 128(%rax)
+; AVX2-FCP-NEXT: vmovdqa %ymm14, 352(%rax)
; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%rax)
; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
@@ -8688,1747 +8544,1675 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FCP-NEXT: vmovaps %ymm0, 416(%rax)
; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT: vmovaps %ymm0, 384(%rax)
-; AVX2-FCP-NEXT: addq $648, %rsp # imm = 0x288
+; AVX2-FCP-NEXT: addq $616, %rsp # imm = 0x268
; AVX2-FCP-NEXT: vzeroupper
; AVX2-FCP-NEXT: retq
;
; AVX512-LABEL: store_i8_stride7_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $1448, %rsp # imm = 0x5A8
-; AVX512-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512-NEXT: vmovdqa %ymm1, %ymm10
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: subq $1720, %rsp # imm = 0x6B8
+; AVX512-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512-NEXT: vpshufb %ymm2, %ymm7, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm26
; AVX512-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
-; AVX512-NEXT: vmovdqa %ymm2, %ymm14
-; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512-NEXT: vmovdqa (%rcx), %ymm14
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm28
; AVX512-NEXT: vmovdqa (%rdx), %ymm8
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512-NEXT: vpshufb %ymm2, %ymm8, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512-NEXT: vpshufb %ymm3, %ymm8, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%r8), %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa (%r8), %ymm15
; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512-NEXT: vmovdqa (%r9), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm2, %ymm15, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512-NEXT: vmovdqa (%r9), %ymm2
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm24
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm11
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512-NEXT: vmovdqa %ymm1, %ymm13
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm11[25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm10
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128]
+; AVX512-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm6, %ymm10, %ymm0
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm9
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm9[23],zero,zero,zero,zero,ymm9[26],zero,ymm9[24],zero,zero,zero,zero,ymm9[27],zero,ymm9[25]
; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
-; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,14,128,12,128,14,15,14,15,128,13,128,15,12,13,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
+; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm10[23,u,u,u],zero,ymm10[26],zero,ymm10[24,u,u,u],zero,ymm10[27],zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm5
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128]
+; AVX512-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm1, %ymm5, %ymm2
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0]
; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm5
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm4, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm20
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
+; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm11
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [29,128,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,128,27,0,0,0,128,30,128,28,0,0,0,128,31,128]
+; AVX512-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm11, %ymm2, %ymm12
+; AVX512-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa 32(%rax), %ymm4
-; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm4, %ymm4
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm1
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22],zero,ymm10[20],zero,zero
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqa 32(%rax), %ymm0
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm12 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm6, %ymm7, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm13
+; AVX512-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm7
+; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm8[23],zero,ymm8[21,22,23,26],zero,ymm8[24],zero,ymm8[28,29,26,27]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm8[18,19,20,21],zero,ymm8[19],zero,ymm8[25,26,27,22],zero,ymm8[20],zero
-; AVX512-NEXT: vmovdqa64 %ymm8, %ymm19
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vpshufb %ymm1, %ymm8, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm14, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm14, %ymm22
+; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rax), %ymm1
-; AVX512-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm1[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpermi2d %zmm0, %zmm2, %zmm16
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm3
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm15, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm15, %ymm16
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512-NEXT: vpshufb %ymm11, %ymm15, %ymm7
+; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0
; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm29
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm30
-; AVX512-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512-NEXT: vpshufb %xmm11, %xmm1, %xmm6
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm7
+; AVX512-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512-NEXT: vpshufb %xmm12, %xmm4, %xmm0
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm15 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-NEXT: vpshufb %xmm15, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm21
-; AVX512-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512-NEXT: vpshufb %xmm6, %xmm14, %xmm7
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512-NEXT: vpor %xmm7, %xmm8, %xmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512-NEXT: vmovdqa 32(%rax), %xmm2
-; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpermi2d %zmm0, %zmm2, %zmm3
-; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512-NEXT: vmovdqa 32(%rax), %xmm0
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpermi2d %zmm7, %zmm8, %zmm1
+; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512-NEXT: vpshufb %xmm6, %xmm0, %xmm8
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm28
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm14 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512-NEXT: vpshufb %xmm14, %xmm2, %xmm10
-; AVX512-NEXT: vmovdqa %xmm2, %xmm3
-; AVX512-NEXT: vporq %xmm8, %xmm10, %xmm26
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm1
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-NEXT: vpshufb %xmm8, %xmm0, %xmm7
+; AVX512-NEXT: vmovdqa64 %xmm8, %xmm20
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm30
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm29
+; AVX512-NEXT: vpor %xmm7, %xmm8, %xmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm28, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm7
; AVX512-NEXT: vmovdqa64 %ymm23, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm7, %ymm8
-; AVX512-NEXT: vmovdqa64 %ymm17, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm9, %ymm10
-; AVX512-NEXT: vpor %ymm8, %ymm10, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm5, %ymm8
+; AVX512-NEXT: vpor %ymm7, %ymm8, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm23
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[0,1,14],zero,ymm0[12,13,0,1,14,15],zero,ymm0[3,12,13,2,3,16],zero,ymm0[30,31,28,29,16,17],zero,ymm0[31,18,19,28,29,18],zero
-; AVX512-NEXT: vpor %ymm8, %ymm10, %ymm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm2
-; AVX512-NEXT: vpshufb %ymm2, %ymm13, %ymm8
-; AVX512-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512-NEXT: vpshufb %ymm2, %ymm11, %ymm10
-; AVX512-NEXT: vmovdqa64 %ymm11, %ymm27
-; AVX512-NEXT: vpor %ymm8, %ymm10, %ymm2
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX512-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm21
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm4, %ymm5, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm26, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm10, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm25, %ymm1
+; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm1
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX512-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm7, %ymm10, %ymm0
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [2,3,4,5,128,3,128,5,4,5,6,128,4,128,6,7,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
+; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm4
+; AVX512-NEXT: vpor %ymm0, %ymm4, %ymm0
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm5
+; AVX512-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm24, %ymm4
+; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm27, %ymm4
+; AVX512-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX512-NEXT: vpor %ymm3, %ymm2, %ymm2
; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512-NEXT: vpshufb %xmm12, %xmm2, %xmm5
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm25
-; AVX512-NEXT: vmovdqa (%rdi), %xmm12
-; AVX512-NEXT: vpshufb %xmm15, %xmm12, %xmm9
-; AVX512-NEXT: vporq %xmm5, %xmm9, %xmm22
-; AVX512-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512-NEXT: vmovdqa64 %xmm29, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm13, %xmm7
-; AVX512-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm9, %xmm10
-; AVX512-NEXT: vpor %xmm7, %xmm10, %xmm2
+; AVX512-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512-NEXT: vpshufb %xmm6, %xmm3, %xmm2
+; AVX512-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512-NEXT: vmovdqa64 %xmm18, %xmm3
+; AVX512-NEXT: vpshufb %xmm3, %xmm4, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %xmm2
+; AVX512-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm2
+; AVX512-NEXT: vmovdqa %xmm3, %xmm6
+; AVX512-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512-NEXT: vpshufb %xmm12, %xmm4, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpshufb %xmm6, %xmm2, %xmm6
-; AVX512-NEXT: vmovdqa (%r8), %xmm5
-; AVX512-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpshufb %xmm14, %xmm5, %xmm11
-; AVX512-NEXT: vpor %xmm6, %xmm11, %xmm6
-; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm6
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm5[8],xmm2[8],xmm5[9],xmm2[9],xmm5[10],xmm2[10],xmm5[11],xmm2[11],xmm5[12],xmm2[12],xmm5[13],xmm2[13],xmm5[14],xmm2[14],xmm5[15],xmm2[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm7 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-NEXT: vpshufb %xmm7, %xmm11, %xmm11
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm11[0,1,0,1],zmm6[4,5,6,7]
+; AVX512-NEXT: vmovdqa (%r9), %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm20, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm3, %xmm2
+; AVX512-NEXT: vmovdqa %xmm3, %xmm8
+; AVX512-NEXT: vmovdqa (%r8), %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm4
+; AVX512-NEXT: vmovdqa %xmm3, %xmm9
+; AVX512-NEXT: vpor %xmm2, %xmm4, %xmm2
; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rax), %xmm10
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm11 = xmm10[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm11, %ymm11
-; AVX512-NEXT: vpshufb {{.*#+}} ymm14 = zero,ymm1[13],zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm11, %zmm24
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm5 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero,ymm13[27],zero,ymm13[25]
+; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512-NEXT: vpshufb %ymm7, %ymm3, %ymm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23,u,u,u],zero,ymm3[26],zero,ymm3[24,u,u,u],zero,ymm3[27],zero
+; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm22, %ymm2
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm3
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm5, %ymm2, %ymm11
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm31 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800,18446463693966278655,18446742978476179455,18446463693966278655,18446742978476179455]
-; AVX512-NEXT: vpternlogq $248, %ymm31, %ymm11, %ymm15
-; AVX512-NEXT: vmovdqa64 %xmm28, %xmm6
-; AVX512-NEXT: vmovdqa %xmm3, %xmm8
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm15[0,1,2,3],zmm11[0,1,0,1]
-; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm17
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm11 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm5
-; AVX512-NEXT: vpshufb %ymm11, %ymm5, %ymm15
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm19
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero,ymm0[29],zero,zero,zero
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm11, %ymm0, %ymm1
-; AVX512-NEXT: vpshuflw $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm11 = mem[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm23
-; AVX512-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm15[25],zero,ymm15[23],zero,zero,zero,zero,ymm15[26],zero,ymm15[24],zero,zero
+; AVX512-NEXT: vmovdqa64 %ymm16, %ymm3
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
+; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm10, %ymm15, %ymm1
+; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%rax), %ymm4
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm15
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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]
+; AVX512-NEXT: vmovdqa64 %xmm9, %xmm21
+; AVX512-NEXT: vmovdqa64 %xmm8, %xmm22
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm2, %xmm24
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm1[0,1,0,1],zmm0[4,5,6,7]
+; AVX512-NEXT: vmovdqa (%rax), %xmm11
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
+; AVX512-NEXT: vpshufb %ymm9, %ymm4, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3],xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[2,3,2,3],zmm2[0,1,0,1]
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm13[8],xmm9[8],xmm13[9],xmm9[9],xmm13[10],xmm9[10],xmm13[11],xmm9[11],xmm13[12],xmm9[12],xmm13[13],xmm9[13],xmm13[14],xmm9[14],xmm13[15],xmm9[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-NEXT: vpshufb %xmm0, %xmm15, %xmm3
-; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm26
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-NEXT: vpshufb %xmm0, %xmm13, %xmm13
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm25
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[2,3,2,3],zmm13[0,1,0,1]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
+; AVX512-NEXT: vmovdqa %xmm6, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm18, %xmm7
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512-NEXT: vpshufb %xmm1, %xmm4, %xmm2
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm30
-; AVX512-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm3
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm12[8],xmm3[9],xmm12[9],xmm3[10],xmm12[10],xmm3[11],xmm12[11],xmm3[12],xmm12[12],xmm3[13],xmm12[13],xmm3[14],xmm12[14],xmm3[15],xmm12[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm4 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
+; AVX512-NEXT: vmovdqa64 %xmm31, %xmm6
+; AVX512-NEXT: vmovdqa64 %xmm28, %xmm4
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm14 = 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]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-NEXT: vpshufb %xmm0, %xmm14, %xmm13
+; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm0
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm21
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512-NEXT: vpshufb %xmm7, %xmm0, %xmm0
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm26[0,1,0,1],zmm0[0,1,0,1]
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm1 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm1, %ymm0, %ymm6
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm1, %ymm0, %ymm8
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm0[23],zero,ymm0[21,22,23,26],zero,ymm0[24],zero,ymm0[28,29,26,27]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm0[18,19,20,21],zero,ymm0[19],zero,ymm0[25,26,27,22],zero,ymm0[20],zero
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22],zero,ymm5[20],zero,zero
-; AVX512-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm29, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm1, %ymm4
-; AVX512-NEXT: vmovdqa %ymm2, %ymm0
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
-; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm27
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm14
-; AVX512-NEXT: vmovdqa64 %ymm14, %ymm25
-; AVX512-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm29
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = zero,ymm0[13],zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm14, %ymm14
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3],xmm12[4],xmm3[4],xmm12[5],xmm3[5],xmm12[6],xmm3[6],xmm12[7],xmm3[7]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512-NEXT: vpshufb %xmm14, %xmm15, %xmm15
-; AVX512-NEXT: vpshufb %xmm14, %xmm12, %xmm12
-; AVX512-NEXT: vinserti32x4 $2, %xmm22, %zmm12, %zmm0
-; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm14 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512-NEXT: vpshufb %xmm14, %xmm11, %xmm9
-; AVX512-NEXT: vpshufb %xmm14, %xmm13, %xmm11
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm6[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm22 = ymm4[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm8[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm7[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm5[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm12 # 16-byte Folded Reload
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3],xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm11 = zmm5[0,1,0,1],mem[0,1,0,1]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm5 = xmm10[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm7 = xmm10[1,1,0,0,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,2,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm7
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm10 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm19[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogq $236, %zmm31, %zmm10, %zmm5
-; AVX512-NEXT: vpandq %ymm31, %ymm22, %ymm10
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm10 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm10, %zmm1, %zmm1
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655,18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512-NEXT: vpand %ymm6, %ymm10, %ymm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm6, %zmm0, %zmm0
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm0, %zmm5
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm6 = zmm23[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogq $236, %zmm10, %zmm1, %zmm6
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm17 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm1, %zmm17, %zmm1
-; AVX512-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm1
-; AVX512-NEXT: vpermq {{.*#+}} zmm6 = zmm30[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpermq {{.*#+}} zmm17 = zmm21[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm17
-; AVX512-NEXT: vpternlogq $248, %ymm10, %ymm13, %ymm14
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm3[2,3,2,3]
-; AVX512-NEXT: vpternlogq $236, %ymm10, %ymm4, %ymm0
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm9[0,1,0,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm14, %zmm3
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm4 = mem[2,3,2,3]
-; AVX512-NEXT: vpshufhw $190, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm6 = mem[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm6
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm15[0,1,0,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512-NEXT: vpternlogq $184, %zmm3, %zmm6, %zmm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm28[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm20[2,3,2,3]
-; AVX512-NEXT: vpor %ymm3, %ymm8, %ymm3
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm3
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm13 = mem[2,3,2,3]
-; AVX512-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm14 = mem[1,1,0,0,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[0,1,2,0]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm15 = mem[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm18 = ymm27[2,3,2,3]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm19 = mem[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm20 = ymm25[2,3,2,3]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm21 = mem[2,3,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm22 = ymm29[2,3,2,3]
-; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm6, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm3 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm6 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
+; AVX512-NEXT: vmovdqa64 %xmm24, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm2[0,1,0,1],zmm1[0,1,0,1]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512-NEXT: vpshufb %ymm9, %ymm10, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512-NEXT: vpternlogq $226, %zmm12, %zmm10, %zmm17
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm2 = mem[2,3,2,3]
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm24
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm24
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm14[0,0,1,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm13, %zmm3
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm3
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm28 = mem[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm28, %zmm3, %zmm28
+; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm10, %zmm28
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm9[2,2,3,2]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm10
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512-NEXT: vmovdqa64 %xmm27, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm27 # 16-byte Folded Reload
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
+; AVX512-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512-NEXT: vpshufb %xmm1, %xmm2, %xmm2
+; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
+; AVX512-NEXT: vmovdqa64 %xmm21, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm22, %xmm3
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
+; AVX512-NEXT: vmovdqa64 %xmm25, %xmm1
+; AVX512-NEXT: vpshufb %xmm1, %xmm8, %xmm8
+; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm11[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm11[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm9, %zmm11, %zmm9
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm11, %zmm12, %zmm11
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm11
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm9, %zmm12, %zmm9
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512-NEXT: vpternlogq $184, %zmm11, %zmm12, %zmm9
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm13[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm18[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,3,2,3]
+; AVX512-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm1 = mem[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
+; AVX512-NEXT: vpermq {{.*#+}} zmm13 = zmm19[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm0
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm21 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm13, %zmm21, %zmm13
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm20 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm18 = zmm30[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vporq %zmm20, %zmm18, %zmm18
+; AVX512-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm18
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm4
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm16
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
; AVX512-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
-; AVX512-NEXT: vporq %ymm15, %ymm18, %ymm5
-; AVX512-NEXT: vporq %ymm19, %ymm20, %ymm6
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm6[0,1,2,3],zmm5[0,1,2,3]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm16
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm16
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm26
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm26
-; AVX512-NEXT: vporq %ymm21, %ymm22, %ymm1
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[0,1,2,3],zmm1[0,1,2,3]
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm1
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm1
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm14
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm14
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm10
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm27[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm0 = mem[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm12[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm7[0,0,1,0,4,4,5,4]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm0
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,0,4,4,5,4]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm2 = zmm2[2,3,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm15
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm15
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512-NEXT: vmovdqa64 %zmm15, 128(%rax)
; AVX512-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm26, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm10, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm14, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)
; AVX512-NEXT: vmovdqa64 %zmm4, 384(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm24, 64(%rax)
-; AVX512-NEXT: addq $1448, %rsp # imm = 0x5A8
+; AVX512-NEXT: vmovdqa64 %zmm16, 64(%rax)
+; AVX512-NEXT: addq $1720, %rsp # imm = 0x6B8
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i8_stride7_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $1256, %rsp # imm = 0x4E8
-; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm14
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm13
+; AVX512-FCP-NEXT: subq $1432, %rsp # imm = 0x598
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128,25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128]
+; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm28
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero
-; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm23
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm1
; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm23
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[0,1,14],zero,ymm1[12,13,0,1,14,15],zero,ymm1[3,12,13,2,3,16],zero,ymm1[30,31,28,29,16,17],zero,ymm1[31,18,19,28,29,18],zero
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm4
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm26
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm3
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm25
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm5
-; AVX512-FCP-NEXT: vmovdqu %ymm5, (%rsp) # 32-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512-FCP-NEXT: vporq %ymm3, %ymm5, %ymm24
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm5
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm28
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm19
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
-; AVX512-FCP-NEXT: vpor %xmm5, %xmm6, %xmm3
-; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm6, %xmm5
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm9
-; AVX512-FCP-NEXT: vmovdqa64 %xmm10, %xmm27
-; AVX512-FCP-NEXT: vpor %xmm5, %xmm9, %xmm5
-; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm10
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm15, %xmm9
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm10, %xmm12
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm21
-; AVX512-FCP-NEXT: vporq %xmm9, %xmm12, %xmm22
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm13, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm13, %ymm20
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
-; AVX512-FCP-NEXT: vpor %ymm7, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm7
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero,zero,zero,ymm7[18]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm7
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,1,14],zero,ymm7[12,13,0,1,14,15],zero,ymm7[3,12,13,2,3,16],zero,ymm7[30,31,28,29,16,17],zero,ymm7[31,18,19,28,29,18],zero
-; AVX512-FCP-NEXT: vpor %ymm1, %ymm7, %ymm1
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
-; AVX512-FCP-NEXT: vmovdqu64 %ymm18, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm13
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm13, %xmm0
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm0
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[13,u,u,u,u,u],zero,ymm0[14,u,u,u,u,u],zero,ymm0[15,u,u,u,u,u],zero,ymm0[16,u,u,u,u,u],zero,ymm0[17,u,u,u]
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm9
+; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vpor %xmm7, %xmm9, %xmm7
+; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
+; AVX512-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm11, %xmm11
+; AVX512-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
+; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm9
+; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm11
+; AVX512-FCP-NEXT: vmovdqa64 %xmm14, %xmm30
+; AVX512-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
+; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm7
+; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm1
-; AVX512-FCP-NEXT: vporq %xmm0, %xmm1, %xmm31
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm14, %xmm0
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
-; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm1, %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm4, %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm21
+; AVX512-FCP-NEXT: vpor %xmm7, %xmm10, %xmm7
+; AVX512-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm6
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm8
+; AVX512-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
+; AVX512-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm11
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm13
+; AVX512-FCP-NEXT: vmovdqa %xmm6, %xmm9
+; AVX512-FCP-NEXT: vporq %xmm12, %xmm13, %xmm31
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm4, %ymm12
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm6
+; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm12
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
+; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm20
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm29
+; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm12
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm28
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm13
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm15
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm27
+; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm15
+; AVX512-FCP-NEXT: vpor %ymm13, %ymm15, %ymm6
+; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm15
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm14
+; AVX512-FCP-NEXT: vpor %ymm15, %ymm14, %ymm6
+; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
+; AVX512-FCP-NEXT: vpor %ymm5, %ymm4, %ymm4
+; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
+; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm23
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27,u,u,u],zero,ymm0[30],zero,ymm0[28,u,u,u],zero,ymm0[31],zero
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm0, %ymm22
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,0,1,14],zero,ymm3[14,15,0,1,14,15],zero,ymm3[13,14,15,16,17,16],zero,ymm3[30,31,30,31,16,17],zero,ymm3[31,28,29,30,31]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero,ymm0[27],zero,ymm0[25]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm12
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[21],zero,ymm12[19],zero,zero,zero,zero,ymm12[22],zero,ymm12[20],zero,zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm11[25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22],zero,ymm11[20]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
-; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm30
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm11[8],xmm9[9],xmm11[9],xmm9[10],xmm11[10],xmm9[11],xmm11[11],xmm9[12],xmm11[12],xmm9[13],xmm11[13],xmm9[14],xmm11[14],xmm9[15],xmm11[15]
+; AVX512-FCP-NEXT: vmovdqa %xmm9, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %xmm11, %xmm17
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[4,5,6,7]
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [2,2,3,3,2,2,3,3]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [2,2,3,3,2,2,3,3]
+; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm4
-; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm4, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm24
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
-; AVX512-FCP-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm26, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3],xmm10[4],xmm15[4],xmm10[5],xmm15[5],xmm10[6],xmm15[6],xmm10[7],xmm15[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm2[0,1,2,3],zmm0[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm14[8],xmm8[8],xmm14[9],xmm8[9],xmm14[10],xmm8[10],xmm14[11],xmm8[11],xmm14[12],xmm8[12],xmm14[13],xmm8[13],xmm14[14],xmm8[14],xmm14[15],xmm8[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm28
-; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm6, %xmm2
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm13[8],xmm9[8],xmm13[9],xmm9[9],xmm13[10],xmm9[10],xmm13[11],xmm9[11],xmm13[12],xmm9[12],xmm13[13],xmm9[13],xmm13[14],xmm9[14],xmm13[15],xmm9[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm4, %xmm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm27
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm15[8],xmm10[9],xmm15[9],xmm10[10],xmm15[10],xmm10[11],xmm15[11],xmm10[12],xmm15[12],xmm10[13],xmm15[13],xmm10[14],xmm15[14],xmm10[15],xmm15[15]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm22[0,1,0,1],zmm2[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm0
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm29
-; AVX512-FCP-NEXT: vpermd %ymm2, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm1 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm20 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm4
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm30
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20],zero,zero
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [4,5,4,5,5,7,4,5]
-; AVX512-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm16
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
-; AVX512-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm15
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm15[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm13[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm31, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm7, %xmm7
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm13
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm0[23],zero,ymm0[23,24,25,26],zero,ymm0[24],zero,ymm0[30,31]
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm1
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm12
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm12[13],zero,zero,zero,zero,zero,zero,ymm12[14],zero,zero,zero,zero,zero,zero,ymm12[15],zero,zero,zero,zero,zero,zero,ymm12[16],zero,zero,zero,zero,zero,zero,ymm12[17],zero,zero
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm20
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm0
; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm8[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqa %ymm12, %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm12[30],zero,ymm12[28],zero,zero,zero,zero,ymm12[31],zero,ymm12[29],zero,zero,zero
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm23[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm11[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm31 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm14
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vpor %ymm12, %ymm9, %ymm9
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm9, %ymm5, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm7
-; AVX512-FCP-NEXT: vpor %ymm4, %ymm10, %ymm4
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm2[8],xmm11[9],xmm2[9],xmm11[10],xmm2[10],xmm11[11],xmm2[11],xmm11[12],xmm2[12],xmm11[13],xmm2[13],xmm11[14],xmm2[14],xmm11[15],xmm2[15]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm1
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm4
-; AVX512-FCP-NEXT: vpor %ymm8, %ymm14, %ymm2
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm18
+; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm2
+; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm21
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm14[8],xmm4[8],xmm14[9],xmm4[9],xmm14[10],xmm4[10],xmm14[11],xmm4[11],xmm14[12],xmm4[12],xmm14[13],xmm4[13],xmm14[14],xmm4[14],xmm14[15],xmm4[15]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm5
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm0, %zmm5
-; AVX512-FCP-NEXT: vpandq %ymm9, %ymm22, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm20, %zmm0
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpandq %ymm26, %ymm19, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm4, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512-FCP-NEXT: vpandq %ymm26, %ymm18, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm4, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm2, %zmm4, %zmm0
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm28[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm27[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm8
-; AVX512-FCP-NEXT: vpandq %ymm26, %ymm13, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm2, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm2, %zmm6, %zmm9
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm9
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm18 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm18 = zmm1[0,1,0,1],mem[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,0,1,2,0,0,1]
-; AVX512-FCP-NEXT: vpermd %ymm2, %ymm4, %ymm19
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpermd %ymm6, %ymm4, %ymm17
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm10
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm12
-; AVX512-FCP-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm15
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm2[23],zero,ymm2[23,24,25,26],zero,ymm2[24],zero,ymm2[30,31]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm14
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm3[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [4,5,4,5,5,7,4,5]
-; AVX512-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm20
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm22 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm22
-; AVX512-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm23 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,0]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm13
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm5
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm15[8],xmm5[9],xmm15[9],xmm5[10],xmm15[10],xmm5[11],xmm15[11],xmm5[12],xmm15[12],xmm5[13],xmm15[13],xmm5[14],xmm15[14],xmm5[15],xmm15[15]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm3
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,6]
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm26
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm24 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm2
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm31[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm14, %xmm1
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm30 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm31, %zmm30, %zmm30
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm30
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm31 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm30, %zmm31, %zmm14
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm18[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm13
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm7, %zmm29, %zmm7
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm27 = zmm27[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vporq %zmm28, %zmm27, %zmm27
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm27
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm23[2,3,2,3],zmm7[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3],xmm5[4],xmm15[4],xmm5[5],xmm15[5],xmm5[6],xmm15[6],xmm5[7],xmm15[7]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm10
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[2,3,2,3],zmm10[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,0,1,2,0,0,1]
+; AVX512-FCP-NEXT: vpermd %ymm11, %ymm17, %ymm28
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm17, %ymm17
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [4,5,4,5,5,7,4,5]
+; AVX512-FCP-NEXT: vpermd %ymm2, %ymm23, %ymm2
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpermd %ymm15, %ymm23, %ymm15
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm15
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm7
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm1 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm18 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm3, %zmm18
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm25[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,0]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,0]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm23, %zmm23 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm23
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm24
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm24
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm2, %zmm0
-; AVX512-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm21
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm21
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
-; AVX512-FCP-NEXT: vpor %ymm12, %ymm15, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm7[0,1,2,3],zmm2[0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm16
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm16
-; AVX512-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm31[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm17, %zmm2
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm2
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm2
-; AVX512-FCP-NEXT: vpor %ymm1, %ymm13, %ymm1
-; AVX512-FCP-NEXT: vpor %ymm11, %ymm14, %ymm5
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[0,1,2,3],zmm1[0,1,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm20, %zmm4
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm4
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm18
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm14 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm14
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm20
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm24[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm9
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm5
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm5
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm6 = zmm3[2,3,2,3],mem[2,3,2,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm6 = zmm3[0,1,2,3],mem[2,3,2,3]
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm11
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 192(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm24, 64(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 384(%rax)
-; AVX512-FCP-NEXT: addq $1256, %rsp # imm = 0x4E8
+; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512-FCP-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i8_stride7_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $1448, %rsp # imm = 0x5A8
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512DQ-NEXT: vmovdqa %ymm1, %ymm10
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: subq $1720, %rsp # imm = 0x6B8
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm7, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm26
; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
-; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm14
-; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm14
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm28
; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm8
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm8, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm8, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512DQ-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm15
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm15, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm2
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm24
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm11
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512DQ-NEXT: vmovdqa %ymm1, %ymm13
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm11[25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm10
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128]
+; AVX512DQ-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm10, %ymm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm9
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm9[23],zero,zero,zero,zero,ymm9[26],zero,ymm9[24],zero,zero,zero,zero,ymm9[27],zero,ymm9[25]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,14,128,12,128,14,15,14,15,128,13,128,15,12,13,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm10[23,u,u,u],zero,ymm10[26],zero,ymm10[24,u,u,u],zero,ymm10[27],zero
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
-; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm5
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128]
+; AVX512DQ-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm5, %ymm2
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0]
; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512DQ-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm5
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm4, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm20
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
+; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm11
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [29,128,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,128,27,0,0,0,128,30,128,28,0,0,0,128,31,128]
+; AVX512DQ-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm2, %ymm12
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm4
-; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm4, %ymm4
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm1
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22],zero,ymm10[20],zero,zero
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm12 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm13 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm7, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm13
+; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm7
+; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm8[23],zero,ymm8[21,22,23,26],zero,ymm8[24],zero,ymm8[28,29,26,27]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm8[18,19,20,21],zero,ymm8[19],zero,ymm8[25,26,27,22],zero,ymm8[20],zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm8, %ymm19
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm8, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm14, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm14, %ymm22
+; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm15, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %ymm15, %ymm16
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm15, %ymm7
+; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rax), %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm1[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm2, %zmm16
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm3
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm0
; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm29
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm30
-; AVX512DQ-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm1, %xmm6
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm7
+; AVX512DQ-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm12 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm4, %xmm0
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm15 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm21
-; AVX512DQ-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm14, %xmm7
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512DQ-NEXT: vpor %xmm7, %xmm8, %xmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm2, %zmm3
-; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm0
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm8, %zmm1
+; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm0, %xmm8
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm28
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm14 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm2, %xmm10
-; AVX512DQ-NEXT: vmovdqa %xmm2, %xmm3
-; AVX512DQ-NEXT: vporq %xmm8, %xmm10, %xmm26
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm1
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-NEXT: vpshufb %xmm8, %xmm0, %xmm7
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm20
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm30
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm29
+; AVX512DQ-NEXT: vpor %xmm7, %xmm8, %xmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm28, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm7
; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm7, %ymm8
-; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm9, %ymm10
-; AVX512DQ-NEXT: vpor %ymm8, %ymm10, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm8
+; AVX512DQ-NEXT: vpor %ymm7, %ymm8, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm23
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[0,1,14],zero,ymm0[12,13,0,1,14,15],zero,ymm0[3,12,13,2,3,16],zero,ymm0[30,31,28,29,16,17],zero,ymm0[31,18,19,28,29,18],zero
-; AVX512DQ-NEXT: vpor %ymm8, %ymm10, %ymm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm13, %ymm8
-; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm11, %ymm10
-; AVX512DQ-NEXT: vmovdqa64 %ymm11, %ymm27
-; AVX512DQ-NEXT: vpor %ymm8, %ymm10, %ymm2
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX512DQ-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm21
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm10, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm1
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX512DQ-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm10, %ymm0
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [2,3,4,5,128,3,128,5,4,5,6,128,4,128,6,7,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm4
+; AVX512DQ-NEXT: vpor %ymm0, %ymm4, %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512DQ-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm5
+; AVX512DQ-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm24, %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm27, %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX512DQ-NEXT: vpor %ymm3, %ymm2, %ymm2
; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm2, %xmm5
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm25
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm12
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm12, %xmm9
-; AVX512DQ-NEXT: vporq %xmm5, %xmm9, %xmm22
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm13, %xmm7
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm9, %xmm10
-; AVX512DQ-NEXT: vpor %xmm7, %xmm10, %xmm2
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm3, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm4, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm6
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm4, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm2, %xmm6
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm5
-; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm5, %xmm11
-; AVX512DQ-NEXT: vpor %xmm6, %xmm11, %xmm6
-; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm6
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm5[8],xmm2[8],xmm5[9],xmm2[9],xmm5[10],xmm2[10],xmm5[11],xmm2[11],xmm5[12],xmm2[12],xmm5[13],xmm2[13],xmm5[14],xmm2[14],xmm5[15],xmm2[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm7 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm11, %xmm11
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm11[0,1,0,1],zmm6[4,5,6,7]
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm3, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm8
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm4
+; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm9
+; AVX512DQ-NEXT: vpor %xmm2, %xmm4, %xmm2
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero,ymm13[27],zero,ymm13[25]
+; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm3, %ymm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rax), %xmm10
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm11 = xmm10[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm11, %ymm11
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm14 = zero,ymm1[13],zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm11, %zmm24
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm5 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23,u,u,u],zero,ymm3[26],zero,ymm3[24,u,u,u],zero,ymm3[27],zero
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm5, %ymm2, %ymm11
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[27],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm31 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800,18446463693966278655,18446742978476179455,18446463693966278655,18446742978476179455]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm31, %ymm11, %ymm15
-; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm6
-; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm8
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm15[0,1,2,3],zmm11[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm17
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm11 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm5
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm5, %ymm15
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm19
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm0[30],zero,ymm0[28],zero,zero,zero,zero,ymm0[31],zero,ymm0[29],zero,zero,zero
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm0, %ymm1
-; AVX512DQ-NEXT: vpshuflw $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm11 = mem[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm23
-; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm15[25],zero,ymm15[23],zero,zero,zero,zero,ymm15[26],zero,ymm15[24],zero,zero
+; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm3
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
+; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm15, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%rax), %ymm4
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm15
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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-NEXT: vmovdqa64 %xmm9, %xmm21
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm22
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm24
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm1[0,1,0,1],zmm0[4,5,6,7]
+; AVX512DQ-NEXT: vmovdqa (%rax), %xmm11
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm4, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3],xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[2,3,2,3],zmm2[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm13[8],xmm9[8],xmm13[9],xmm9[9],xmm13[10],xmm9[10],xmm13[11],xmm9[11],xmm13[12],xmm9[12],xmm13[13],xmm9[13],xmm13[14],xmm9[14],xmm13[15],xmm9[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm15, %xmm3
-; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm26
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm13, %xmm13
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm25
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[2,3,2,3],zmm13[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
+; AVX512DQ-NEXT: vmovdqa %xmm6, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm7
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm4, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm30
-; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm3
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm12[8],xmm3[9],xmm12[9],xmm3[10],xmm12[10],xmm3[11],xmm12[11],xmm3[12],xmm12[12],xmm3[13],xmm12[13],xmm3[14],xmm12[14],xmm3[15],xmm12[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm6
+; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm4
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm14 = 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-NEXT: vmovdqa {{.*#+}} xmm0 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm14, %xmm13
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm0
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm21
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm0, %xmm0
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm26[0,1,0,1],zmm0[0,1,0,1]
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm1 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm0, %ymm6
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm0, %ymm8
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm0[23],zero,ymm0[21,22,23,26],zero,ymm0[24],zero,ymm0[28,29,26,27]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm0[18,19,20,21],zero,ymm0[19],zero,ymm0[25,26,27,22],zero,ymm0[20],zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22],zero,ymm5[20],zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm29, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm1, %ymm4
-; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm0
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
-; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm27
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm14
-; AVX512DQ-NEXT: vmovdqa64 %ymm14, %ymm25
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm29
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = zero,ymm0[13],zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm14, %ymm14
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3],xmm12[4],xmm3[4],xmm12[5],xmm3[5],xmm12[6],xmm3[6],xmm12[7],xmm3[7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm15, %xmm15
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm12, %xmm12
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm22, %zmm12, %zmm0
-; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm14 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm11, %xmm9
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm13, %xmm11
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm6[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm22 = ymm4[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm8[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm7[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm5[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm12 # 16-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm5 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3],xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm11 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm11 = zmm5[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm5 = xmm10[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm7 = xmm10[1,1,0,0,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,2,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm7
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm10 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm19[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogq $236, %zmm31, %zmm10, %zmm5
-; AVX512DQ-NEXT: vpandq %ymm31, %ymm22, %ymm10
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm10 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm10, %zmm1, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm10 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655,18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512DQ-NEXT: vpand %ymm6, %ymm10, %ymm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm6, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm0, %zmm5
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm6 = zmm23[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogq $236, %zmm10, %zmm1, %zmm6
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm17 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm1, %zmm17, %zmm1
-; AVX512DQ-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm6 = zmm30[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm17 = zmm21[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm17
-; AVX512DQ-NEXT: vpternlogq $248, %ymm10, %ymm13, %ymm14
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm3[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $236, %ymm10, %ymm4, %ymm0
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm9[0,1,0,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm14, %zmm3
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm4 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpshufhw $190, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm6 = mem[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm6
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm15[0,1,0,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512DQ-NEXT: vpternlogq $184, %zmm3, %zmm6, %zmm4
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm28[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm20[2,3,2,3]
-; AVX512DQ-NEXT: vpor %ymm3, %ymm8, %ymm3
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm3
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512DQ-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm13 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm14 = mem[1,1,0,0,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[0,1,2,0]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm15 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm27[2,3,2,3]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm19 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm20 = ymm25[2,3,2,3]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm21 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm22 = ymm29[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm6, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm3 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm6 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm2[0,1,0,1],zmm1[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm10, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm12, %zmm10, %zmm17
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm2 = mem[2,3,2,3]
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm24
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm24
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm14[0,0,1,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm13, %zmm3
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm3
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm28 = mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm28, %zmm3, %zmm28
+; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm10, %zmm28
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm9[2,2,3,2]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm10
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm27 # 16-byte Folded Reload
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm2, %xmm2
+; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm22, %xmm3
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm1
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm8, %xmm8
+; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm11[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm11[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm9, %zmm11, %zmm9
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm11, %zmm12, %zmm11
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm11
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm9, %zmm12, %zmm9
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm11, %zmm12, %zmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm13[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm18[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm1 = mem[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm13 = zmm19[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm0
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm21 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm13, %zmm21, %zmm13
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm20 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm18 = zmm30[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vporq %zmm20, %zmm18, %zmm18
+; AVX512DQ-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm18
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm4
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm16
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
; AVX512DQ-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
-; AVX512DQ-NEXT: vporq %ymm15, %ymm18, %ymm5
-; AVX512DQ-NEXT: vporq %ymm19, %ymm20, %ymm6
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm6[0,1,2,3],zmm5[0,1,2,3]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm16
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm16
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm26
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm26
-; AVX512DQ-NEXT: vporq %ymm21, %ymm22, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[0,1,2,3],zmm1[0,1,2,3]
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm1
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm1
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm14
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm14
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm10
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm27[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512DQ-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm0 = mem[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm12[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm7[0,0,1,0,4,4,5,4]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm0
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,0,4,4,5,4]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm2 = zmm2[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm15
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm15
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-NEXT: vmovdqa64 %zmm15, 128(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm26, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm10, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm14, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm4, 384(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm24, 64(%rax)
-; AVX512DQ-NEXT: addq $1448, %rsp # imm = 0x5A8
+; AVX512DQ-NEXT: vmovdqa64 %zmm16, 64(%rax)
+; AVX512DQ-NEXT: addq $1720, %rsp # imm = 0x6B8
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i8_stride7_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $1256, %rsp # imm = 0x4E8
-; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm13
+; AVX512DQ-FCP-NEXT: subq $1432, %rsp # imm = 0x598
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128,25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128]
+; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm28
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[25],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm23
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm22
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm23
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[0,1,14],zero,ymm1[12,13,0,1,14,15],zero,ymm1[3,12,13,2,3,16],zero,ymm1[30,31,28,29,16,17],zero,ymm1[31,18,19,28,29,18],zero
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm4
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm26
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm3
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm25
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm5, (%rsp) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512DQ-FCP-NEXT: vporq %ymm3, %ymm5, %ymm24
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm28
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm19
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
-; AVX512DQ-FCP-NEXT: vpor %xmm5, %xmm6, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm6, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm10, %xmm27
-; AVX512DQ-FCP-NEXT: vpor %xmm5, %xmm9, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm15, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm10, %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm21
-; AVX512DQ-FCP-NEXT: vporq %xmm9, %xmm12, %xmm22
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm13, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm13, %ymm20
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
-; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero,zero,zero,ymm7[18]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[0,1,14],zero,ymm7[12,13,0,1,14,15],zero,ymm7[3,12,13,2,3,16],zero,ymm7[30,31,28,29,16,17],zero,ymm7[31,18,19,28,29,18],zero
-; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm7, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %ymm18, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm13
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm13, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[13,u,u,u,u,u],zero,ymm0[14,u,u,u,u,u],zero,ymm0[15,u,u,u,u,u],zero,ymm0[16,u,u,u,u,u],zero,ymm0[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm9, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm11, %xmm11
+; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm14, %xmm30
+; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm1
-; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm1, %xmm31
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm14, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
-; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm4, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm21
+; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm10, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm8
+; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm11
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm6
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, %xmm9
+; AVX512DQ-FCP-NEXT: vporq %xmm12, %xmm13, %xmm31
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm4, %ymm12
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
+; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm12
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm20
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm29
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm12
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm28
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm13
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm15
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm15
+; AVX512DQ-FCP-NEXT: vpor %ymm13, %ymm15, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm14
+; AVX512DQ-FCP-NEXT: vpor %ymm15, %ymm14, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
+; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
+; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27,u,u,u],zero,ymm0[30],zero,ymm0[28,u,u,u],zero,ymm0[31],zero
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm0, %ymm22
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,0,1,14],zero,ymm3[14,15,0,1,14,15],zero,ymm3[13,14,15,16,17,16],zero,ymm3[30,31,30,31,16,17],zero,ymm3[31,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero,ymm0[27],zero,ymm0[25]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[21],zero,ymm12[19],zero,zero,zero,zero,ymm12[22],zero,ymm12[20],zero,zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm11[25],zero,ymm11[23],zero,zero,zero,zero,ymm11[26],zero,ymm11[24],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22],zero,ymm11[20]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
-; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm30
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm11[8],xmm9[9],xmm11[9],xmm9[10],xmm11[10],xmm9[11],xmm11[11],xmm9[12],xmm11[12],xmm9[13],xmm11[13],xmm9[14],xmm11[14],xmm9[15],xmm11[15]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm11, %xmm17
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[4,5,6,7]
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [2,2,3,3,2,2,3,3]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [2,2,3,3,2,2,3,3]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm24
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
-; AVX512DQ-FCP-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512DQ-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm26, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3],xmm10[4],xmm15[4],xmm10[5],xmm15[5],xmm10[6],xmm15[6],xmm10[7],xmm15[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm2[0,1,2,3],zmm0[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm14[8],xmm8[8],xmm14[9],xmm8[9],xmm14[10],xmm8[10],xmm14[11],xmm8[11],xmm14[12],xmm8[12],xmm14[13],xmm8[13],xmm14[14],xmm8[14],xmm14[15],xmm8[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm28
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm13[8],xmm9[8],xmm13[9],xmm9[9],xmm13[10],xmm9[10],xmm13[11],xmm9[11],xmm13[12],xmm9[12],xmm13[13],xmm9[13],xmm13[14],xmm9[14],xmm13[15],xmm9[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm4, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm27
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm15[8],xmm10[9],xmm15[9],xmm10[10],xmm15[10],xmm10[11],xmm15[11],xmm10[12],xmm15[12],xmm10[13],xmm15[13],xmm10[14],xmm15[14],xmm10[15],xmm15[15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm22[0,1,0,1],zmm2[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm29
-; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm1 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm20 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm4
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm30
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [4,5,4,5,5,7,4,5]
-; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm16
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
-; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm15
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm15[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm18 = ymm13[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm31, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm0[23],zero,ymm0[23,24,25,26],zero,ymm0[24],zero,ymm0[30,31]
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
+; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm1
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm12
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm12[13],zero,zero,zero,zero,zero,zero,ymm12[14],zero,zero,zero,zero,zero,zero,ymm12[15],zero,zero,zero,zero,zero,zero,ymm12[16],zero,zero,zero,zero,zero,zero,ymm12[17],zero,zero
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm20
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm8[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm12, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm12[30],zero,ymm12[28],zero,zero,zero,zero,ymm12[31],zero,ymm12[29],zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm23[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm11[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm31 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm14
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpor %ymm12, %ymm9, %ymm9
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm9, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm10, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm2[8],xmm11[9],xmm2[9],xmm11[10],xmm2[10],xmm11[11],xmm2[11],xmm11[12],xmm2[12],xmm11[13],xmm2[13],xmm11[14],xmm2[14],xmm11[15],xmm2[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm4
-; AVX512DQ-FCP-NEXT: vpor %ymm8, %ymm14, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm18
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm21
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm14[8],xmm4[8],xmm14[9],xmm4[9],xmm14[10],xmm4[10],xmm14[11],xmm4[11],xmm14[12],xmm4[12],xmm14[13],xmm4[13],xmm14[14],xmm4[14],xmm14[15],xmm4[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT: vpandq %ymm9, %ymm22, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm20, %zmm0
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm2, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpandq %ymm26, %ymm19, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm4, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512DQ-FCP-NEXT: vpandq %ymm26, %ymm18, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm4, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm2, %zmm4, %zmm0
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm28[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm27[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm8
-; AVX512DQ-FCP-NEXT: vpandq %ymm26, %ymm13, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm2, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm2, %zmm6, %zmm9
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm9
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm18 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm18 = zmm1[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm3
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,1,0,1,2,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm4, %ymm19
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm6, %ymm4, %ymm17
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm10
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm12
-; AVX512DQ-FCP-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512DQ-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm15
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm2[23],zero,ymm2[23,24,25,26],zero,ymm2[24],zero,ymm2[30,31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm3[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [4,5,4,5,5,7,4,5]
-; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm2, %ymm20
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm22 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm22
-; AVX512DQ-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm23 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,0]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm5
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm15[8],xmm5[9],xmm15[9],xmm5[10],xmm15[10],xmm5[11],xmm15[11],xmm5[12],xmm15[12],xmm5[13],xmm15[13],xmm5[14],xmm15[14],xmm5[15],xmm15[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm3
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,6]
+; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm26
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm24 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm31[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm14, %xmm1
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm30 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm31, %zmm30, %zmm30
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm30
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm31 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm30, %zmm31, %zmm14
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm18[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm13
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm29, %zmm7
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm27 = zmm27[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm28, %zmm27, %zmm27
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm23[2,3,2,3],zmm7[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3],xmm5[4],xmm15[4],xmm5[5],xmm15[5],xmm5[6],xmm15[6],xmm5[7],xmm15[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[2,3,2,3],zmm10[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm2
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,0,1,2,0,0,1]
+; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm17, %ymm28
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm17, %ymm17
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [4,5,4,5,5,7,4,5]
+; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermd %ymm15, %ymm23, %ymm15
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm15
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm7
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm18 = mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm3, %zmm18
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm25[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,0]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,0]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm23, %zmm23 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm23
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm24
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm24
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm21
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm21
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
-; AVX512DQ-FCP-NEXT: vpor %ymm12, %ymm15, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm7[0,1,2,3],zmm2[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm16
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm16
-; AVX512DQ-FCP-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm31[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm17, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm2
-; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm13, %ymm1
-; AVX512DQ-FCP-NEXT: vpor %ymm11, %ymm14, %ymm5
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[0,1,2,3],zmm1[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm20, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm18
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm1
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm14 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm14
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm20
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm24[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm9
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm6 = zmm3[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm6 = zmm3[0,1,2,3],mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm11
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, 64(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 384(%rax)
-; AVX512DQ-FCP-NEXT: addq $1256, %rsp # imm = 0x4E8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512DQ-FCP-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -10436,8 +10220,8 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa (%rax), %ymm13
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512BW-NEXT: vpshufb %ymm27, %ymm13, %ymm0
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-NEXT: vpshufb %ymm26, %ymm13, %ymm0
; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512BW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512BW-NEXT: vpermw %ymm13, %ymm1, %ymm1
@@ -10446,12 +10230,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
; AVX512BW-NEXT: vpshufb %ymm17, %ymm9, %ymm1
; AVX512BW-NEXT: vmovdqa (%r8), %ymm10
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm22 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512BW-NEXT: vpshufb %ymm22, %ymm10, %ymm2
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512BW-NEXT: vpshufb %ymm21, %ymm10, %ymm2
; AVX512BW-NEXT: vpor %ymm1, %ymm2, %ymm2
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm12
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm3
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm3[8],xmm12[8],xmm3[9],xmm12[9],xmm3[10],xmm12[10],xmm3[11],xmm12[11],xmm3[12],xmm12[12],xmm3[13],xmm12[13],xmm3[14],xmm12[14],xmm3[15],xmm12[15]
+; AVX512BW-NEXT: vmovdqa (%r9), %xmm1
+; AVX512BW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm12
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm1[8],xmm12[9],xmm1[9],xmm12[10],xmm1[10],xmm12[11],xmm1[11],xmm12[12],xmm1[12],xmm12[13],xmm1[13],xmm12[14],xmm1[14],xmm12[15],xmm1[15]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm8
@@ -10461,270 +10246,257 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqa (%rdx), %ymm14
; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm0 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
; AVX512BW-NEXT: vpshufb %ymm0, %ymm14, %ymm2
-; AVX512BW-NEXT: vmovdqa64 (%rcx), %ymm16
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512BW-NEXT: vpshufb %ymm23, %ymm16, %ymm4
+; AVX512BW-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512BW-NEXT: vpshufb %ymm20, %ymm15, %ymm4
; AVX512BW-NEXT: vpor %ymm2, %ymm4, %ymm2
; AVX512BW-NEXT: vmovdqa (%rdx), %xmm4
; AVX512BW-NEXT: vmovdqa (%rcx), %xmm5
; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm22
; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm18
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
; AVX512BW-NEXT: vpshufb %ymm24, %ymm18, %ymm2
; AVX512BW-NEXT: vmovdqa64 (%rsi), %ymm19
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512BW-NEXT: vpshufb %ymm26, %ymm19, %ymm6
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512BW-NEXT: vpshufb %ymm25, %ymm19, %ymm6
; AVX512BW-NEXT: vpor %ymm2, %ymm6, %ymm2
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm6
; AVX512BW-NEXT: vmovdqa (%rsi), %xmm7
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm21[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm21, %zmm2
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm23 = xmm23[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,1,0,1]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm23, %zmm3
; AVX512BW-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
; AVX512BW-NEXT: kmovq %r10, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm20, %zmm2 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm22, %zmm3 {%k1}
; AVX512BW-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
; AVX512BW-NEXT: kmovq %r10, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm2 {%k1}
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm3 {%k1}
; AVX512BW-NEXT: vmovdqa64 32(%rdx), %ymm29
+; AVX512BW-NEXT: vpshufb %ymm0, %ymm29, %ymm0
+; AVX512BW-NEXT: vmovdqa64 32(%rcx), %ymm30
+; AVX512BW-NEXT: vpshufb %ymm20, %ymm30, %ymm8
+; AVX512BW-NEXT: vpor %ymm0, %ymm8, %ymm0
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
; AVX512BW-NEXT: vpshufb %ymm20, %ymm29, %ymm8
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512BW-NEXT: vpshufb %ymm22, %ymm30, %ymm23
+; AVX512BW-NEXT: vporq %ymm8, %ymm23, %ymm8
; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm21 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512BW-NEXT: vmovdqa64 32(%rcx), %ymm30
-; AVX512BW-NEXT: vpshufb %ymm21, %ymm30, %ymm25
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512BW-NEXT: vporq %ymm8, %ymm25, %ymm8
-; AVX512BW-NEXT: vpshufb %ymm0, %ymm29, %ymm0
-; AVX512BW-NEXT: vpshufb %ymm23, %ymm30, %ymm23
-; AVX512BW-NEXT: vporq %ymm0, %ymm23, %ymm0
; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
; AVX512BW-NEXT: vmovdqa64 32(%rsi), %ymm28
-; AVX512BW-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT: vmovdqa64 32(%rdi), %ymm16
+; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm16[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
-; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm25 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
+; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm23 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
; AVX512BW-NEXT: movl $676341840, %r10d # imm = 0x28502850
; AVX512BW-NEXT: kmovd %r10d, %k1
-; AVX512BW-NEXT: vpshufb %ymm25, %ymm28, %ymm8 {%k1}
+; AVX512BW-NEXT: vpshufb %ymm23, %ymm28, %ymm8 {%k1}
; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512BW-NEXT: vpshufb %ymm24, %ymm15, %ymm23
-; AVX512BW-NEXT: vpshufb %ymm26, %ymm28, %ymm24
-; AVX512BW-NEXT: vporq %ymm23, %ymm24, %ymm23
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm23, %zmm8
+; AVX512BW-NEXT: vpshufb %ymm24, %ymm16, %ymm24
+; AVX512BW-NEXT: vpshufb %ymm25, %ymm28, %ymm25
+; AVX512BW-NEXT: vporq %ymm24, %ymm25, %ymm24
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm24, %zmm8
; AVX512BW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
; AVX512BW-NEXT: kmovq %r10, %k2
; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm8 {%k2}
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm23 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512BW-NEXT: vmovdqa64 32(%r8), %ymm31
-; AVX512BW-NEXT: vpshufb %ymm23, %ymm31, %ymm0
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm26 = ymm0[2,3,2,3]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512BW-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512BW-NEXT: vpshufb %ymm24, %ymm0, %ymm11
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512BW-NEXT: vporq %ymm26, %ymm11, %ymm11
-; AVX512BW-NEXT: vpshufb %ymm17, %ymm0, %ymm17
-; AVX512BW-NEXT: vpshufb %ymm22, %ymm31, %ymm22
-; AVX512BW-NEXT: vporq %ymm17, %ymm22, %ymm17
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm17, %zmm11
+; AVX512BW-NEXT: vmovdqa64 32(%r9), %ymm31
+; AVX512BW-NEXT: vpshufb %ymm17, %ymm31, %ymm17
+; AVX512BW-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512BW-NEXT: vpshufb %ymm21, %ymm1, %ymm21
+; AVX512BW-NEXT: vporq %ymm17, %ymm21, %ymm17
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512BW-NEXT: vpshufb %ymm24, %ymm1, %ymm21
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm25 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512BW-NEXT: vpshufb %ymm25, %ymm31, %ymm27
+; AVX512BW-NEXT: vporq %ymm21, %ymm27, %ymm21
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,3,2,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm17, %zmm21
; AVX512BW-NEXT: vmovdqa64 32(%rax), %ymm17
-; AVX512BW-NEXT: vpshufb %ymm27, %ymm17, %ymm22
+; AVX512BW-NEXT: vpshufb %ymm26, %ymm17, %ymm27
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm26 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-NEXT: vpermw %ymm17, %ymm26, %ymm27
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm27, %zmm22, %zmm22
+; AVX512BW-NEXT: vpermw %ymm17, %ymm26, %ymm11
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm27, %zmm11
; AVX512BW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
; AVX512BW-NEXT: kmovq %r10, %k3
-; AVX512BW-NEXT: vmovdqu8 %zmm22, %zmm11 {%k3}
+; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm21 {%k3}
; AVX512BW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
; AVX512BW-NEXT: kmovq %r10, %k3
-; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm8 {%k3}
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm15[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT: vmovdqu8 %zmm21, %zmm8 {%k3}
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm16[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[2,2,3,3,6,6,7,7]
; AVX512BW-NEXT: movl $338170920, %r10d # imm = 0x14281428
; AVX512BW-NEXT: kmovd %r10d, %k4
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm27 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
; AVX512BW-NEXT: vpshufb %ymm27, %ymm28, %ymm11 {%k4}
-; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512BW-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512BW-NEXT: vpshufb %ymm1, %ymm28, %ymm22
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512BW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512BW-NEXT: vpshufb %ymm2, %ymm28, %ymm21
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm28 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512BW-NEXT: vpshufb %ymm28, %ymm15, %ymm15
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512BW-NEXT: vporq %ymm22, %ymm15, %ymm15
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm15
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm29[24,25],zero,ymm29[23],zero,ymm29[21,22,23,26],zero,ymm29[24],zero,ymm29[28,29,26,27]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm22 = ymm30[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm30[25],zero,ymm30[23],zero,zero,zero,zero,ymm30[26],zero,ymm30[24],zero,zero,zero,zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
-; AVX512BW-NEXT: vporq %ymm11, %ymm22, %ymm22
+; AVX512BW-NEXT: vpshufb %ymm28, %ymm16, %ymm16
+; AVX512BW-NEXT: vporq %ymm21, %ymm16, %ymm16
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm16
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm29[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm11[0,2,3,3,4,6,7,7]
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm11 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512BW-NEXT: vpshufb %ymm11, %ymm30, %ymm30
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm29 = ymm29[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm29 = ymm29[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT: vmovdqu8 %ymm29, %ymm30 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm29 = ymm30[2,3,2,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm29, %zmm22, %zmm22
+; AVX512BW-NEXT: vpshufb %ymm11, %ymm30, %ymm0
+; AVX512BW-NEXT: vmovdqu8 %ymm21, %ymm0 {%k1}
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm21 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm29[24,25],zero,ymm29[23],zero,ymm29[21,22,23,26],zero,ymm29[24],zero,ymm29[28,29,26,27]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm29 = ymm30[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm30[25],zero,ymm30[23],zero,zero,zero,zero,ymm30[26],zero,ymm30[24],zero,zero,zero,zero
+; AVX512BW-NEXT: vporq %ymm21, %ymm29, %ymm21
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
; AVX512BW-NEXT: kmovq %r10, %k3
-; AVX512BW-NEXT: vmovdqu8 %zmm22, %zmm15 {%k3}
-; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm29
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm31[0,1,2,3],zmm29[4,5,6,7]
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm22[23],zero,zmm22[23,24,25,26],zero,zmm22[24],zero,zmm22[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm30
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm30[4,5,6,7]
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT: vporq %zmm22, %zmm0, %zmm0
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm0
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm29
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm31[0,1,2,3],zmm29[4,5,6,7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm21 = zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm21[25],zero,zmm21[23],zero,zero,zero,zero,zmm21[26],zero,zmm21[24],zero,zero,zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm21[59],zero,zero,zero,zero,zmm21[62],zero,zmm21[60],zero,zero,zero,zero,zmm21[63],zero,zmm21[61]
+; AVX512BW-NEXT: vporq %zmm1, %zmm21, %zmm1
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
; AVX512BW-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
; AVX512BW-NEXT: kmovq %r10, %k3
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm15 {%k3}
-; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm22
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512BW-NEXT: vpermi2w %zmm22, %zmm17, %zmm0
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm16 {%k3}
+; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm21
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512BW-NEXT: vpermi2w %zmm21, %zmm17, %zmm1
; AVX512BW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512BW-NEXT: kmovq %rax, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm15 {%k5}
-; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,1,1,4,4,5,5]
-; AVX512BW-NEXT: vpshufb %ymm25, %ymm19, %ymm0 {%k1}
-; AVX512BW-NEXT: vpshufb %ymm1, %ymm19, %ymm1
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-NEXT: vpshufb %ymm28, %ymm18, %ymm25
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512BW-NEXT: vporq %ymm1, %ymm25, %ymm1
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm1
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
-; AVX512BW-NEXT: vpshufb %zmm20, %zmm1, %zmm1
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm16 {%k5}
+; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,1,1,4,4,5,5]
+; AVX512BW-NEXT: vpshufb %ymm23, %ymm19, %ymm1 {%k1}
+; AVX512BW-NEXT: vpshufb %ymm2, %ymm19, %ymm2
+; AVX512BW-NEXT: vpshufb %ymm28, %ymm18, %ymm23
+; AVX512BW-NEXT: vporq %ymm2, %ymm23, %ymm2
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm2
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512BW-NEXT: vpshufb %zmm20, %zmm2, %zmm2
; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm20
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm20
-; AVX512BW-NEXT: vpshufb %zmm21, %zmm20, %zmm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm15, %zmm20, %zmm20
+; AVX512BW-NEXT: vpshufb %zmm22, %zmm20, %zmm20
+; AVX512BW-NEXT: vporq %zmm2, %zmm20, %zmm2
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT: vporq %zmm1, %zmm20, %zmm20
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm20 {%k3}
-; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm21
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm10, %zmm30, %zmm0
-; AVX512BW-NEXT: vpshufb %zmm23, %zmm0, %zmm0
-; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm23
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm9, %zmm29, %zmm1
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm10, %zmm29, %zmm1
; AVX512BW-NEXT: vpshufb %zmm24, %zmm1, %zmm1
-; AVX512BW-NEXT: vmovdqa64 32(%rdx), %xmm24
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm23
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512BW-NEXT: vpshufb %zmm25, %zmm0, %zmm2
+; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512BW-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512BW-NEXT: vmovdqa 32(%rcx), %xmm2
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-NEXT: vporq %zmm0, %zmm1, %zmm1
-; AVX512BW-NEXT: vmovdqa64 32(%rcx), %xmm25
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm13, %zmm22, %zmm13
-; AVX512BW-NEXT: vpermw %zmm13, %zmm26, %zmm0
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm13, %zmm21, %zmm13
+; AVX512BW-NEXT: vpermw %zmm13, %zmm26, %zmm24
; AVX512BW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
; AVX512BW-NEXT: kmovq %rax, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k5}
-; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512BW-NEXT: vmovdqu8 %zmm24, %zmm1 {%k5}
+; AVX512BW-NEXT: vmovdqa64 32(%rdi), %xmm24
; AVX512BW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
; AVX512BW-NEXT: kmovq %rax, %k5
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm20 {%k5}
; AVX512BW-NEXT: vmovdqa 32(%rsi), %xmm1
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm18 = ymm18[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm26 = ymm18[2,2,3,3,6,6,7,7]
-; AVX512BW-NEXT: vpshufb %ymm27, %ymm19, %ymm26 {%k4}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm18[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT: vpshufb %ymm27, %ymm19, %ymm25 {%k4}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm24[0],xmm1[0],xmm24[1],xmm1[1],xmm24[2],xmm1[2],xmm24[3],xmm1[3],xmm24[4],xmm1[4],xmm24[5],xmm1[5],xmm24[6],xmm1[6],xmm24[7],xmm1[7]
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm18 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
; AVX512BW-NEXT: vpshufb %xmm18, %xmm19, %xmm19
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm26[2,3,2,3],zmm19[0,1,0,1]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm24[0],xmm25[0],xmm24[1],xmm25[1],xmm24[2],xmm25[2],xmm24[3],xmm25[3],xmm24[4],xmm25[4],xmm24[5],xmm25[5],xmm24[6],xmm25[6],xmm24[7],xmm25[7]
-; AVX512BW-NEXT: vpshufb %ymm11, %ymm16, %ymm11
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm16 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-NEXT: vpshufb %xmm16, %xmm26, %xmm26
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm25[2,3,2,3],zmm19[0,1,0,1]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512BW-NEXT: vpshufb %ymm11, %ymm15, %ymm11
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm19 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm26, %xmm15
; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
; AVX512BW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,2,3,3,4,6,7,7]
; AVX512BW-NEXT: vmovdqu8 %ymm14, %ymm11 {%k1}
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm11[2,3,2,3],zmm26[0,1,0,1]
-; AVX512BW-NEXT: vmovdqu8 %zmm19, %zmm14 {%k2}
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm11[2,3,2,3],zmm15[0,1,0,1]
+; AVX512BW-NEXT: vmovdqu8 %zmm25, %zmm14 {%k2}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm15 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-NEXT: vpshufb %xmm15, %xmm11, %xmm11
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm9[27],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero,ymm10[29]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512BW-NEXT: vpor %ymm9, %ymm10, %ymm10
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm23[0],xmm21[0],xmm23[1],xmm21[1],xmm23[2],xmm21[2],xmm23[3],xmm21[3],xmm23[4],xmm21[4],xmm23[5],xmm21[5],xmm23[6],xmm21[6],xmm23[7],xmm21[7]
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm11 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512BW-NEXT: vpermw %zmm22, %zmm11, %zmm11
+; AVX512BW-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm9[2,3,2,3],zmm11[0,1,0,1]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512BW-NEXT: vpermw %zmm21, %zmm10, %zmm10
; AVX512BW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm10 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm9 {%k1}
; AVX512BW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm10, %zmm14 {%k1}
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm25, %xmm10
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm24, %xmm26
-; AVX512BW-NEXT: vporq %xmm10, %xmm26, %xmm10
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm24 = xmm24[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm24, %zmm10, %zmm10
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm24 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm1, %xmm25
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm26 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512BW-NEXT: vpshufb %xmm26, %xmm0, %xmm27
-; AVX512BW-NEXT: vporq %xmm25, %xmm27, %xmm25
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm25, %zmm0
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3}
+; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm14 {%k1}
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm2, %xmm9
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm0, %xmm25
+; AVX512BW-NEXT: vporq %xmm9, %xmm25, %xmm9
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm9, %zmm0
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm2 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm9
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm24, %xmm26
+; AVX512BW-NEXT: vporq %xmm9, %xmm26, %xmm9
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm24[8],xmm1[9],xmm24[9],xmm1[10],xmm24[10],xmm1[11],xmm24[11],xmm1[12],xmm24[12],xmm1[13],xmm24[13],xmm1[14],xmm24[14],xmm1[15],xmm24[15]
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm9, %zmm1
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm9 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm9 {%k3}
; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm0 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm21, %xmm1
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512BW-NEXT: vpermi2w %zmm22, %zmm17, %zmm25
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm22, %xmm1
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm24 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512BW-NEXT: vpermi2w %zmm21, %zmm17, %zmm24
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm17 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-NEXT: vpshufb %xmm17, %xmm23, %xmm22
-; AVX512BW-NEXT: vporq %xmm1, %xmm22, %xmm1
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm23[8],xmm21[8],xmm23[9],xmm21[9],xmm23[10],xmm21[10],xmm23[11],xmm21[11],xmm23[12],xmm21[12],xmm23[13],xmm21[13],xmm23[14],xmm21[14],xmm23[15],xmm21[15]
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm23, %xmm21
+; AVX512BW-NEXT: vporq %xmm1, %xmm21, %xmm1
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm21[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512BW-NEXT: vinserti32x4 $2, %xmm21, %zmm1, %zmm1
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
; AVX512BW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm25, %zmm1 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm24, %zmm1 {%k1}
; AVX512BW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k1}
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm5, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm4, %xmm11
-; AVX512BW-NEXT: vpor %xmm1, %xmm11, %xmm1
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm9 {%k1}
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm5, %xmm1
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm4, %xmm10
+; AVX512BW-NEXT: vpor %xmm1, %xmm10, %xmm1
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512BW-NEXT: vpshufb %xmm16, %xmm4, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm4, %xmm4
; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm4, %zmm1
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm7, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm26, %xmm6, %xmm5
-; AVX512BW-NEXT: vpor %xmm4, %xmm5, %xmm4
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512BW-NEXT: vpshufb %xmm18, %xmm5, %xmm5
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm7, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm6, %xmm4
+; AVX512BW-NEXT: vpor %xmm2, %xmm4, %xmm2
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512BW-NEXT: vpshufb %xmm18, %xmm4, %xmm4
+; AVX512BW-NEXT: vinserti32x4 $2, %xmm2, %zmm4, %zmm2
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
; AVX512BW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm4 {%k1}
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm12, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm17, %xmm3, %xmm1
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512BW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm4, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm12, %xmm1
; AVX512BW-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
+; AVX512BW-NEXT: vpshufb %xmm15, %xmm1, %xmm1
; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
; AVX512BW-NEXT: vpermw %zmm13, %zmm1, %zmm1
@@ -10734,334 +10506,319 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
; AVX512BW-NEXT: kmovq %rax, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k1}
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm8, 320(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm10, 256(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm9, 256(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm14, 192(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm20, 128(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm15, 384(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm16, 384(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BW-FCP-LABEL: store_i8_stride7_vf64:
; AVX512BW-FCP: # %bb.0:
-; AVX512BW-FCP-NEXT: subq $40, %rsp
+; AVX512BW-FCP-NEXT: subq $104, %rsp
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT: vmovdqa (%rax), %ymm0
-; AVX512BW-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm28 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512BW-FCP-NEXT: vpshufb %ymm28, %ymm0, %ymm1
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpermw %ymm0, %ymm2, %ymm2
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm2
-; AVX512BW-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512BW-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %ymm20
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm22 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm22, %ymm20, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa (%rax), %ymm2
+; AVX512BW-FCP-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm2, %ymm0
+; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpermw %ymm2, %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm2
+; AVX512BW-FCP-NEXT: vmovdqa (%r9), %ymm1
+; AVX512BW-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm3
+; AVX512BW-FCP-NEXT: vmovdqa (%r8), %ymm4
+; AVX512BW-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm4
; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm0
-; AVX512BW-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm4
-; AVX512BW-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %xmm16
+; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm15
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm15[8],xmm16[8],xmm15[9],xmm16[9],xmm15[10],xmm16[10],xmm15[11],xmm16[11],xmm15[12],xmm16[12],xmm15[13],xmm16[13],xmm15[14],xmm16[14],xmm15[15],xmm16[15]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512BW-FCP-NEXT: movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
; AVX512BW-FCP-NEXT: kmovq %r10, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm2
-; AVX512BW-FCP-NEXT: vmovdqa %ymm0, %ymm9
-; AVX512BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm5
-; AVX512BW-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm5, %ymm2
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm14
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm15
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm5
-; AVX512BW-FCP-NEXT: vmovdqa %ymm0, %ymm6
-; AVX512BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm24
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm21 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512BW-FCP-NEXT: vpshufb %ymm21, %ymm24, %ymm16
-; AVX512BW-FCP-NEXT: vporq %ymm5, %ymm16, %ymm5
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm17
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm25 = xmm25[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm25, %zmm13
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512BW-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512BW-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm6
+; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm17
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm19
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm19[8],xmm17[8],xmm19[9],xmm17[9],xmm19[10],xmm17[10],xmm19[11],xmm17[11],xmm19[12],xmm17[12],xmm19[13],xmm17[13],xmm19[14],xmm17[14],xmm19[15],xmm17[15]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm29 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm29, %xmm6, %xmm6
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm6
+; AVX512BW-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512BW-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512BW-FCP-NEXT: vpshufb %ymm23, %ymm7, %ymm21
+; AVX512BW-FCP-NEXT: vporq %ymm6, %ymm21, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm21
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm22
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm24, %xmm24
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm24, %zmm14
; AVX512BW-FCP-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm13 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm14 {%k1}
; AVX512BW-FCP-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm13 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm3
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm3[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm26
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm4
-; AVX512BW-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm11
-; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm11, %ymm4
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm4, %zmm4
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm25
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm27
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm27, %ymm26
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512BW-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm12
-; AVX512BW-FCP-NEXT: vpshufb %ymm21, %ymm27, %ymm21
-; AVX512BW-FCP-NEXT: vporq %ymm12, %ymm21, %ymm12
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm12, %zmm12
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm14 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm3
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm5
+; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm3, %ymm5
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm24
+; AVX512BW-FCP-NEXT: vporq %ymm5, %ymm24, %ymm5
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm9
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm25
+; AVX512BW-FCP-NEXT: vpshufb %ymm23, %ymm25, %ymm23
+; AVX512BW-FCP-NEXT: vporq %ymm9, %ymm23, %ymm9
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm23
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm25, %ymm24
+; AVX512BW-FCP-NEXT: vporq %ymm23, %ymm24, %ymm23
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm9, %zmm9
; AVX512BW-FCP-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm12 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %ymm4
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm21
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm21[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm26
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512BW-FCP-NEXT: vpshufb %ymm22, %ymm4, %ymm22
-; AVX512BW-FCP-NEXT: vporq %ymm1, %ymm22, %ymm1
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm1, %zmm1
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rax), %ymm31
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm22 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
-; AVX512BW-FCP-NEXT: vpermw %ymm31, %ymm22, %ymm22
-; AVX512BW-FCP-NEXT: vpshufb %ymm28, %ymm31, %ymm25
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm25, %zmm22
-; AVX512BW-FCP-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm9 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm28
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm28, %ymm1
+; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm28, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm23
+; AVX512BW-FCP-NEXT: vporq %ymm1, %ymm23, %ymm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rax), %ymm6
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512BW-FCP-NEXT: vpermw %ymm6, %ymm23, %ymm23
+; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm6, %ymm20
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512BW-FCP-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm1 {%k2}
; AVX512BW-FCP-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm12 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm25
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm25[4,5,6,7]
+; AVX512BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm9 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm23
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm23[4,5,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,zmm1[23],zero,zmm1[21,22,23,26],zero,zmm1[24],zero,zmm1[28,29,26,27,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm26
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm26[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm24
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm24[4,5,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm28
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm27[0,1,2,3],zmm28[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm26
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm25[0,1,2,3],zmm26[4,5,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[27],zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm2[60],zero,zmm2[62,63,62,63],zero,zmm2[61],zero,zmm2[63,60,61]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm29
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm11[0,1,2,3],zmm29[4,5,6,7]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm27
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm5[0,1,2,3],zmm27[4,5,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[23],zero,zero,zero,zero,zmm3[26],zero,zmm3[24],zero,zero,zero,zero,zmm3[27],zero,zmm3[25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zmm3[62],zero,zmm3[60],zero,zero,zero,zero,zmm3[63],zero,zmm3[61],zero,zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vporq %zmm2, %zmm3, %zmm22
+; AVX512BW-FCP-NEXT: vporq %zmm2, %zmm3, %zmm2
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
-; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm22 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm27
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[0,1,2,3],zmm27[4,5,6,7]
+; AVX512BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm25
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm28[0,1,2,3],zmm25[4,5,6,7]
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm28
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm28[4,5,6,7]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm2[59],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61]
+; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm30
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm30[4,5,6,7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512BW-FCP-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
-; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm22 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm10
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512BW-FCP-NEXT: vpermi2w %zmm10, %zmm31, %zmm0
+; AVX512BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm31
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512BW-FCP-NEXT: vpermi2w %zmm31, %zmm6, %zmm1
; AVX512BW-FCP-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm22 {%k3}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm6[28],zero,ymm6[30,31,30,31],zero,ymm6[29],zero,ymm6[31,28,29]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm24[30],zero,ymm24[28],zero,zero,zero,zero,ymm24[31],zero,ymm24[29],zero,zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm8
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm6
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm18[30],zero,ymm18[28],zero,zero,zero,zero,ymm18[31],zero,ymm18[29],zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm4, %ymm4
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm18 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm0, %xmm0
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm9
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k2}
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm19[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm19[27],zero,zero,zero,zero,ymm19[30],zero,ymm19[28],zero,zero,zero,zero,ymm19[31],zero
-; AVX512BW-FCP-NEXT: vmovdqa64 %ymm19, %ymm21
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm20[27],zero,zero,zero,zero,ymm20[30],zero,ymm20[28],zero,zero,zero,zero,ymm20[31],zero,ymm20[29]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm6, %ymm3
-; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %xmm5
-; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm0, %xmm0
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512BW-FCP-NEXT: vpermw %zmm10, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k2}
-; AVX512BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm9 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm3
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm23
-; AVX512BW-FCP-NEXT: vporq %xmm3, %xmm23, %xmm3
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm1
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm3
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm23
-; AVX512BW-FCP-NEXT: vporq %xmm3, %xmm23, %xmm3
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm7[8],xmm8[8],xmm7[9],xmm8[9],xmm7[10],xmm8[10],xmm7[11],xmm8[11],xmm7[12],xmm8[12],xmm7[13],xmm8[13],xmm7[14],xmm8[14],xmm7[15],xmm8[15]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm23 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm23 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm8
-; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm8, %xmm3
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm4, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512BW-FCP-NEXT: vpermi2w %zmm10, %zmm31, %zmm4
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm1
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm5, %xmm2
+; AVX512BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
+; AVX512BW-FCP-NEXT: vpshufb %xmm29, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm1, %zmm3
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm29
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm30
+; AVX512BW-FCP-NEXT: vporq %xmm29, %xmm30, %xmm29
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm30 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
+; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm30, %xmm30
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm30, %zmm29, %zmm29
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm29 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm30
+; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm30, %xmm0
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm18
+; AVX512BW-FCP-NEXT: vporq %xmm0, %xmm18, %xmm0
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm3[8],xmm30[8],xmm3[9],xmm30[9],xmm3[10],xmm30[10],xmm3[11],xmm30[11],xmm3[12],xmm30[12],xmm3[13],xmm30[13],xmm3[14],xmm30[14],xmm3[15],xmm30[15]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm18 = xmm18[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm18, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm18 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512BW-FCP-NEXT: vpermi2w %zmm31, %zmm6, %zmm18
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
; AVX512BW-FCP-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k2}
+; AVX512BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm0 {%k3}
; AVX512BW-FCP-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm23 {%k2}
-; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm3
-; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm14, %xmm4
-; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3],xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm18, %xmm4, %xmm4
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm4, %zmm3
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm17, %xmm2
-; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm16, %xmm0
-; AVX512BW-FCP-NEXT: vpor %xmm2, %xmm0, %xmm0
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm16[0],xmm17[0],xmm16[1],xmm17[1],xmm16[2],xmm17[2],xmm16[3],xmm17[3],xmm16[4],xmm17[4],xmm16[5],xmm17[5],xmm16[6],xmm17[6],xmm16[7],xmm17[7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm0
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm29 {%k3}
+; AVX512BW-FCP-NEXT: vpshufb %xmm13, %xmm19, %xmm0
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm17, %xmm6
+; AVX512BW-FCP-NEXT: vpor %xmm0, %xmm6, %xmm0
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm0
+; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm22, %xmm6
+; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm21, %xmm8
+; AVX512BW-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm21[0],xmm22[0],xmm21[1],xmm22[1],xmm21[2],xmm22[2],xmm21[3],xmm22[3],xmm21[4],xmm22[4],xmm21[5],xmm22[5],xmm21[6],xmm22[6],xmm21[7],xmm22[7]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm8
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm8, %zmm6
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm6[0,1,0,1,4,5,4,5]
; AVX512BW-FCP-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm0 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512BW-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512BW-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm2
-; AVX512BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm2, %xmm2
-; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm10, %zmm2 # 32-byte Folded Reload
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-FCP-NEXT: vpermw %zmm2, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm8, %zmm0 {%k3}
+; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm16, %xmm6
+; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm15, %xmm7
+; AVX512BW-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm7
+; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm7, %zmm6
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm31, %zmm8 # 32-byte Folded Reload
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm7 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
+; AVX512BW-FCP-NEXT: vpermw %zmm8, %zmm7, %zmm7
; AVX512BW-FCP-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm1 {%k2}
+; AVX512BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k3}
; AVX512BW-FCP-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k2}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm24, %zmm28, %zmm1
+; AVX512BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm0 {%k3}
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm1
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm13[28],zero,ymm13[30,31,30,31],zero,ymm13[29],zero,ymm13[31,28,29]
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm6, %ymm2
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[2,3,2,3],zmm2[0,1,0,1]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm30[0],xmm3[1],xmm30[1],xmm3[2],xmm30[2],xmm3[3],xmm30[3],xmm3[4],xmm30[4],xmm3[5],xmm30[5],xmm3[6],xmm30[6],xmm3[7],xmm30[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
+; AVX512BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[2,3,2,3],zmm1[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512BW-FCP-NEXT: vpermw %zmm31, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm26, %zmm1
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,19,20,21],zero,zmm1[19],zero,zmm1[21,20,21,22],zero,zmm1[20],zero,zmm1[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero,zero,zero,zmm1[59],zero
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm3 # 32-byte Folded Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm27, %zmm3
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero,zmm3[59],zero,zmm3[57]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm3, %zmm1
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm3 # 32-byte Folded Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm24, %zmm3
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[18],zero,zmm3[18,19,20,21],zero,zmm3[19],zero,zmm3[25,26,27,22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[56,57],zero,zmm3[55],zero,zmm3[53,54,55,58],zero,zmm3[56],zero,zmm3[60,61,58,59]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm4 # 32-byte Folded Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm23, %zmm4
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22],zero,zmm4[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero,zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm3, %zmm4, %zmm3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k1}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm20, %zmm30, %zmm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm1
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm27, %zmm4
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm25, %zmm4
; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[20],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm4, %zmm1
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-FCP-NEXT: vpermw %zmm2, %zmm4, %zmm2
+; AVX512BW-FCP-NEXT: vpermw %zmm8, %zmm4, %zmm4
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
; AVX512BW-FCP-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm1 {%k1}
; AVX512BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k1}
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 320(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, 256(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 384(%rax)
-; AVX512BW-FCP-NEXT: addq $40, %rsp
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 256(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 384(%rax)
+; AVX512BW-FCP-NEXT: addq $104, %rsp
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
@@ -11069,8 +10826,8 @@ define void @store_i8_stride7_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: vmovdqa (%rax), %ymm13
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm13, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm13, %ymm0
; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
; AVX512DQ-BW-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-BW-NEXT: vpermw %ymm13, %ymm1, %ymm1
@@ -11079,12 +10836,13 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm9, %ymm1
; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm10
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm22 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-BW-NEXT: vpshufb %ymm22, %ymm10, %ymm2
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512DQ-BW-NEXT: vpshufb %ymm21, %ymm10, %ymm2
; AVX512DQ-BW-NEXT: vpor %ymm1, %ymm2, %ymm2
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm12
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm3
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm3[8],xmm12[8],xmm3[9],xmm12[9],xmm3[10],xmm12[10],xmm3[11],xmm12[11],xmm3[12],xmm12[12],xmm3[13],xmm12[13],xmm3[14],xmm12[14],xmm3[15],xmm12[15]
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm12
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm1[8],xmm12[9],xmm1[9],xmm12[10],xmm1[10],xmm12[11],xmm1[11],xmm12[12],xmm1[12],xmm12[13],xmm1[13],xmm12[14],xmm1[14],xmm12[15],xmm1[15]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm8
@@ -11094,270 +10852,257 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm14
; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm0 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm14, %ymm2
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %ymm16
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm16, %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm15, %ymm4
; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm4, %ymm2
; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm4
; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm5
; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm22
; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %ymm18
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm18, %ymm2
; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %ymm19
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm19, %ymm6
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm19, %ymm6
; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm6, %ymm2
; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm6
; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm7
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm21[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm21, %zmm2
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm23 = xmm23[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,1,0,1]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm23, %zmm3
; AVX512DQ-BW-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
; AVX512DQ-BW-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm20, %zmm2 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm22, %zmm3 {%k1}
; AVX512DQ-BW-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
; AVX512DQ-BW-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm2 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm3 {%k1}
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %ymm29
+; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm29, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %ymm30
+; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm30, %ymm8
+; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm8, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm29, %ymm8
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm22 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-BW-NEXT: vpshufb %ymm22, %ymm30, %ymm23
+; AVX512DQ-BW-NEXT: vporq %ymm8, %ymm23, %ymm8
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm21 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %ymm30
-; AVX512DQ-BW-NEXT: vpshufb %ymm21, %ymm30, %ymm25
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512DQ-BW-NEXT: vporq %ymm8, %ymm25, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm29, %ymm0
-; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm30, %ymm23
-; AVX512DQ-BW-NEXT: vporq %ymm0, %ymm23, %ymm0
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %ymm28
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %ymm16
+; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm16[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-NEXT: vpbroadcastd {{.*#+}} ymm25 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
+; AVX512DQ-BW-NEXT: vpbroadcastd {{.*#+}} ymm23 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
; AVX512DQ-BW-NEXT: movl $676341840, %r10d # imm = 0x28502850
; AVX512DQ-BW-NEXT: kmovd %r10d, %k1
-; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm28, %ymm8 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm28, %ymm8 {%k1}
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm15, %ymm23
-; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm28, %ymm24
-; AVX512DQ-BW-NEXT: vporq %ymm23, %ymm24, %ymm23
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm23, %zmm8
+; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm16, %ymm24
+; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm28, %ymm25
+; AVX512DQ-BW-NEXT: vporq %ymm24, %ymm25, %ymm24
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm24, %zmm8
; AVX512DQ-BW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
; AVX512DQ-BW-NEXT: kmovq %r10, %k2
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm8 {%k2}
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm23 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %ymm31
-; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm31, %ymm0
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm26 = ymm0[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm0, %ymm11
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512DQ-BW-NEXT: vporq %ymm26, %ymm11, %ymm11
-; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm0, %ymm17
-; AVX512DQ-BW-NEXT: vpshufb %ymm22, %ymm31, %ymm22
-; AVX512DQ-BW-NEXT: vporq %ymm17, %ymm22, %ymm17
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm17, %zmm11
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %ymm31
+; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm31, %ymm17
+; AVX512DQ-BW-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512DQ-BW-NEXT: vpshufb %ymm21, %ymm1, %ymm21
+; AVX512DQ-BW-NEXT: vporq %ymm17, %ymm21, %ymm17
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm1, %ymm21
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm25 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm31, %ymm27
+; AVX512DQ-BW-NEXT: vporq %ymm21, %ymm27, %ymm21
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,3,2,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm17, %zmm21
; AVX512DQ-BW-NEXT: vmovdqa64 32(%rax), %ymm17
-; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm17, %ymm22
+; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm17, %ymm27
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm26 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-NEXT: vpermw %ymm17, %ymm26, %ymm27
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm27, %zmm22, %zmm22
+; AVX512DQ-BW-NEXT: vpermw %ymm17, %ymm26, %ymm11
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm27, %zmm11
; AVX512DQ-BW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
; AVX512DQ-BW-NEXT: kmovq %r10, %k3
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm22, %zmm11 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm21 {%k3}
; AVX512DQ-BW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
; AVX512DQ-BW-NEXT: kmovq %r10, %k3
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm8 {%k3}
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm15[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm21, %zmm8 {%k3}
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm16[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[2,2,3,3,6,6,7,7]
; AVX512DQ-BW-NEXT: movl $338170920, %r10d # imm = 0x14281428
; AVX512DQ-BW-NEXT: kmovd %r10d, %k4
; AVX512DQ-BW-NEXT: vpbroadcastd {{.*#+}} ymm27 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm28, %ymm11 {%k4}
-; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512DQ-BW-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpshufb %ymm1, %ymm28, %ymm22
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512DQ-BW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpshufb %ymm2, %ymm28, %ymm21
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm28 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm15, %ymm15
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-BW-NEXT: vporq %ymm22, %ymm15, %ymm15
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm15
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm11 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm29[24,25],zero,ymm29[23],zero,ymm29[21,22,23,26],zero,ymm29[24],zero,ymm29[28,29,26,27]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm22 = ymm30[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm30[25],zero,ymm30[23],zero,zero,zero,zero,ymm30[26],zero,ymm30[24],zero,zero,zero,zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
-; AVX512DQ-BW-NEXT: vporq %ymm11, %ymm22, %ymm22
+; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm16, %ymm16
+; AVX512DQ-BW-NEXT: vporq %ymm21, %ymm16, %ymm16
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm16
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm29[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm11[0,2,3,3,4,6,7,7]
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm11 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm30, %ymm30
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm29 = ymm29[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm29 = ymm29[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm29, %ymm30 {%k1}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm29 = ymm30[2,3,2,3]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm29, %zmm22, %zmm22
+; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm30, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm21, %ymm0 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm21 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm29[24,25],zero,ymm29[23],zero,ymm29[21,22,23,26],zero,ymm29[24],zero,ymm29[28,29,26,27]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm29 = ymm30[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm30[25],zero,ymm30[23],zero,zero,zero,zero,ymm30[26],zero,ymm30[24],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vporq %ymm21, %ymm29, %ymm21
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm16 = zmm16[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
; AVX512DQ-BW-NEXT: kmovq %r10, %k3
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm22, %zmm15 {%k3}
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm29
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm31[0,1,2,3],zmm29[4,5,6,7]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm22[23],zero,zmm22[23,24,25,26],zero,zmm22[24],zero,zmm22[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm30
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm30[4,5,6,7]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT: vporq %zmm22, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm16 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm0
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm29
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm31[0,1,2,3],zmm29[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm21 = zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm21[25],zero,zmm21[23],zero,zero,zero,zero,zmm21[26],zero,zmm21[24],zero,zero,zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm21[59],zero,zero,zero,zero,zmm21[62],zero,zmm21[60],zero,zero,zero,zero,zmm21[63],zero,zmm21[61]
+; AVX512DQ-BW-NEXT: vporq %zmm1, %zmm21, %zmm1
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
; AVX512DQ-BW-NEXT: kmovq %r10, %k3
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm15 {%k3}
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rax), %zmm22
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512DQ-BW-NEXT: vpermi2w %zmm22, %zmm17, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm16 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rax), %zmm21
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512DQ-BW-NEXT: vpermi2w %zmm21, %zmm17, %zmm1
; AVX512DQ-BW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512DQ-BW-NEXT: kmovq %rax, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm15 {%k5}
-; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm19, %ymm0 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %ymm1, %ymm19, %ymm1
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm18, %ymm25
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512DQ-BW-NEXT: vporq %ymm1, %ymm25, %ymm1
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm1
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpshufb %zmm20, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm16 {%k5}
+; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm19, %ymm1 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb %ymm2, %ymm19, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm18, %ymm23
+; AVX512DQ-BW-NEXT: vporq %ymm2, %ymm23, %ymm2
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm2
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: vpshufb %zmm20, %zmm2, %zmm2
; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm20
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm20
-; AVX512DQ-BW-NEXT: vpshufb %zmm21, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm15, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT: vpshufb %zmm22, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT: vporq %zmm2, %zmm20, %zmm2
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT: vporq %zmm1, %zmm20, %zmm20
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm20 {%k3}
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm21
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm10, %zmm30, %zmm0
-; AVX512DQ-BW-NEXT: vpshufb %zmm23, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm23
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm9, %zmm29, %zmm1
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm10, %zmm29, %zmm1
; AVX512DQ-BW-NEXT: vpshufb %zmm24, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %xmm24
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm23
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpshufb %zmm25, %zmm0, %zmm2
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512DQ-BW-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rcx), %xmm2
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-NEXT: vporq %zmm0, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %xmm25
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm13, %zmm22, %zmm13
-; AVX512DQ-BW-NEXT: vpermw %zmm13, %zmm26, %zmm0
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm13, %zmm21, %zmm13
+; AVX512DQ-BW-NEXT: vpermw %zmm13, %zmm26, %zmm24
; AVX512DQ-BW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
; AVX512DQ-BW-NEXT: kmovq %rax, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k5}
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm24, %zmm1 {%k5}
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %xmm24
; AVX512DQ-BW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
; AVX512DQ-BW-NEXT: kmovq %rax, %k5
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm20 {%k5}
; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %xmm1
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm18 = ymm18[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm26 = ymm18[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm19, %ymm26 {%k4}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm18[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT: vpshufb %ymm27, %ymm19, %ymm25 {%k4}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm24[0],xmm1[0],xmm24[1],xmm1[1],xmm24[2],xmm1[2],xmm24[3],xmm1[3],xmm24[4],xmm1[4],xmm24[5],xmm1[5],xmm24[6],xmm1[6],xmm24[7],xmm1[7]
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm18 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
; AVX512DQ-BW-NEXT: vpshufb %xmm18, %xmm19, %xmm19
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm26[2,3,2,3],zmm19[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm24[0],xmm25[0],xmm24[1],xmm25[1],xmm24[2],xmm25[2],xmm24[3],xmm25[3],xmm24[4],xmm25[4],xmm24[5],xmm25[5],xmm24[6],xmm25[6],xmm24[7],xmm25[7]
-; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm16, %ymm11
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm16 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm16, %xmm26, %xmm26
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm25[2,3,2,3],zmm19[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm15, %ymm11
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm19 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm26, %xmm15
; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,2,3,3,4,6,7,7]
; AVX512DQ-BW-NEXT: vmovdqu8 %ymm14, %ymm11 {%k1}
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm11[2,3,2,3],zmm26[0,1,0,1]
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm19, %zmm14 {%k2}
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm11[2,3,2,3],zmm15[0,1,0,1]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm25, %zmm14 {%k2}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm15 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm11, %xmm11
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm9[27],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero,ymm10[29]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm10, %ymm10
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm23[0],xmm21[0],xmm23[1],xmm21[1],xmm23[2],xmm21[2],xmm23[3],xmm21[3],xmm23[4],xmm21[4],xmm23[5],xmm21[5],xmm23[6],xmm21[6],xmm23[7],xmm21[7]
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm11 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512DQ-BW-NEXT: vpermw %zmm22, %zmm11, %zmm11
+; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm9[2,3,2,3],zmm11[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512DQ-BW-NEXT: vpermw %zmm21, %zmm10, %zmm10
; AVX512DQ-BW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm10 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm9 {%k1}
; AVX512DQ-BW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm10, %zmm14 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm25, %xmm10
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm24, %xmm26
-; AVX512DQ-BW-NEXT: vporq %xmm10, %xmm26, %xmm10
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm24 = xmm24[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm24, %zmm10, %zmm10
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm24 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm1, %xmm25
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm26 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm0, %xmm27
-; AVX512DQ-BW-NEXT: vporq %xmm25, %xmm27, %xmm25
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm25, %zmm0
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm14 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm2, %xmm9
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm0, %xmm25
+; AVX512DQ-BW-NEXT: vporq %xmm9, %xmm25, %xmm9
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm9, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm2 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm1, %xmm9
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm24, %xmm26
+; AVX512DQ-BW-NEXT: vporq %xmm9, %xmm26, %xmm9
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm24[8],xmm1[9],xmm24[9],xmm1[10],xmm24[10],xmm1[11],xmm24[11],xmm1[12],xmm24[12],xmm1[13],xmm24[13],xmm1[14],xmm24[14],xmm1[15],xmm24[15]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm9, %zmm1
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm9 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm9 {%k3}
; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm0 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512DQ-BW-NEXT: vpermi2w %zmm22, %zmm17, %zmm25
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm22, %xmm1
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm24 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512DQ-BW-NEXT: vpermi2w %zmm21, %zmm17, %zmm24
; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm17 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm23, %xmm22
-; AVX512DQ-BW-NEXT: vporq %xmm1, %xmm22, %xmm1
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm23[8],xmm21[8],xmm23[9],xmm21[9],xmm23[10],xmm21[10],xmm23[11],xmm21[11],xmm23[12],xmm21[12],xmm23[13],xmm21[13],xmm23[14],xmm21[14],xmm23[15],xmm21[15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm23, %xmm21
+; AVX512DQ-BW-NEXT: vporq %xmm1, %xmm21, %xmm1
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm21 = xmm21[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm21, %zmm1, %zmm1
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
; AVX512DQ-BW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm25, %zmm1 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm24, %zmm1 {%k1}
; AVX512DQ-BW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm5, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm4, %xmm11
-; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm11, %xmm1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm9 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm5, %xmm1
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm4, %xmm10
+; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm10, %xmm1
; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm16, %xmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm4, %xmm4
; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm4, %zmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm7, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm6, %xmm5
-; AVX512DQ-BW-NEXT: vpor %xmm4, %xmm5, %xmm4
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm18, %xmm5, %xmm5
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm7, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm6, %xmm4
+; AVX512DQ-BW-NEXT: vpor %xmm2, %xmm4, %xmm2
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm18, %xmm4, %xmm4
+; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm2, %zmm4, %zmm2
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
; AVX512DQ-BW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm4 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm12, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm3, %xmm1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm4, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm12, %xmm1
; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm1, %xmm1
; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
; AVX512DQ-BW-NEXT: vpermw %zmm13, %zmm1, %zmm1
@@ -11367,334 +11112,319 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512DQ-BW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 320(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 256(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 256(%rax)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 192(%rax)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 128(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 384(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 384(%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
; AVX512DQ-BW-FCP-LABEL: store_i8_stride7_vf64:
; AVX512DQ-BW-FCP: # %bb.0:
-; AVX512DQ-BW-FCP-NEXT: subq $40, %rsp
+; AVX512DQ-BW-FCP-NEXT: subq $104, %rsp
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rax), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm28 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm28, %ymm0, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512DQ-BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm0, %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %ymm20
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm22 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm22, %ymm20, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rax), %ymm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm2, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQ-BW-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm2, %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm4
; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %xmm16
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm15
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm15[8],xmm16[8],xmm15[9],xmm16[9],xmm15[10],xmm16[10],xmm15[11],xmm16[11],xmm15[12],xmm16[12],xmm15[13],xmm16[13],xmm15[14],xmm16[14],xmm15[15],xmm16[15]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512DQ-BW-FCP-NEXT: movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, %ymm9
-; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm5, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm14
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm15
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm24
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm21 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm21, %ymm24, %ymm16
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm5, %ymm16, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm17
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm25 = xmm25[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm25, %zmm13
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm6, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm17
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm19
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm19[8],xmm17[8],xmm19[9],xmm17[9],xmm19[10],xmm17[10],xmm19[11],xmm17[11],xmm19[12],xmm17[12],xmm19[13],xmm17[13],xmm19[14],xmm17[14],xmm19[15],xmm17[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm29 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm29, %xmm6, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm6, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-BW-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm23, %ymm7, %ymm21
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm6, %ymm21, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm21
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm22
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm24, %xmm24
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm24, %zmm14
; AVX512DQ-BW-FCP-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm13 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm14 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm13 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm3[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm26
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm11
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm11, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm4, %zmm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm27
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm27, %ymm26
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm12
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm21, %ymm27, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm12, %ymm21, %ymm12
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm12, %zmm12
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm14 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm5
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm3, %ymm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm24
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm5, %ymm24, %ymm5
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm25
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm23, %ymm25, %ymm23
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm9, %ymm23, %ymm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm23
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm25, %ymm24
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm23, %ymm24, %ymm23
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm9, %zmm9
; AVX512DQ-BW-FCP-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm12 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %ymm4
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm25 = ymm21[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm26
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm25, %ymm26, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm22, %ymm4, %ymm22
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm1, %ymm22, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm1, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rax), %ymm31
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm22 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm31, %ymm22, %ymm22
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm28, %ymm31, %ymm25
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm25, %zmm22
-; AVX512DQ-BW-FCP-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm9 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm28
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm28, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm28, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm23
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm1, %ymm23, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rax), %ymm6
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm6, %ymm23, %ymm23
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm6, %ymm20
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512DQ-BW-FCP-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm1 {%k2}
; AVX512DQ-BW-FCP-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm12 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm25
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm25[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm9 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm23
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm23[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,zmm1[23],zero,zmm1[21,22,23,26],zero,zmm1[24],zero,zmm1[28,29,26,27,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero,zmm1[61],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm26
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm26[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm24
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm24[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm28
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm27[0,1,2,3],zmm28[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm26
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm25[0,1,2,3],zmm26[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zero,zero,zmm2[27],zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,60,61,62],zero,zmm2[60],zero,zmm2[62,63,62,63],zero,zmm2[61],zero,zmm2[63,60,61]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm29
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm11[0,1,2,3],zmm29[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm27
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm5[0,1,2,3],zmm27[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[23],zero,zero,zero,zero,zmm3[26],zero,zmm3[24],zero,zero,zero,zero,zmm3[27],zero,zmm3[25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zmm3[62],zero,zmm3[60],zero,zero,zero,zero,zmm3[63],zero,zmm3[61],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vporq %zmm2, %zmm3, %zmm22
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm2, %zmm3, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm2[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm22 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm27
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[0,1,2,3],zmm27[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm25
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm28[0,1,2,3],zmm25[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm28
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm28[4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm2[25],zero,zmm2[23],zero,zero,zero,zero,zmm2[26],zero,zmm2[24],zero,zero,zmm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm2[59],zero,zero,zero,zero,zmm2[62],zero,zmm2[60],zero,zero,zero,zero,zmm2[63],zero,zmm2[61]
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm30
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm30[4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[25],zero,zmm0[23],zero,zero,zero,zero,zmm0[26],zero,zmm0[24],zero,zero,zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm0[59],zero,zero,zero,zero,zmm0[62],zero,zmm0[60],zero,zero,zero,zero,zmm0[63],zero,zmm0[61]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm22 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm10
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm10, %zmm31, %zmm0
+; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm31
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm31, %zmm6, %zmm1
; AVX512DQ-BW-FCP-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm22 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm6[28],zero,ymm6[30,31,30,31],zero,ymm6[29],zero,ymm6[31,28,29]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm24[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm24[30],zero,ymm24[28],zero,zero,zero,zero,ymm24[31],zero,ymm24[29],zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm8
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm18[30],zero,ymm18[28],zero,zero,zero,zero,ymm18[31],zero,ymm18[29],zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm9[30],zero,ymm9[28],zero,zero,zero,zero,ymm9[31],zero,ymm9[29],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm4, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm18 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm0, %xmm0
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm9
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm9 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm19[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm19[27],zero,zero,zero,zero,ymm19[30],zero,ymm19[28],zero,zero,zero,zero,ymm19[31],zero
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %ymm19, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm20[27],zero,zero,zero,zero,ymm20[30],zero,ymm20[28],zero,zero,zero,zero,ymm20[31],zero,ymm20[29]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm6, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %xmm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm0, %xmm0
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm10, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k2}
-; AVX512DQ-BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm9 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm23
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm3, %xmm23, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm23
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm3, %xmm23, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm7[8],xmm8[8],xmm7[9],xmm8[9],xmm7[10],xmm8[10],xmm7[11],xmm8[11],xmm7[12],xmm8[12],xmm7[13],xmm8[13],xmm7[14],xmm8[14],xmm7[15],xmm8[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm7, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm23 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm23 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm8
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm8, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm4, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm10, %zmm31, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm5, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm29, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm1, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm29
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm30
+; AVX512DQ-BW-FCP-NEXT: vporq %xmm29, %xmm30, %xmm29
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm30 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm30, %xmm30
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm30, %zmm29, %zmm29
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm29 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm30
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm30, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm18
+; AVX512DQ-BW-FCP-NEXT: vporq %xmm0, %xmm18, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm3[8],xmm30[8],xmm3[9],xmm30[9],xmm3[10],xmm30[10],xmm3[11],xmm30[11],xmm3[12],xmm30[12],xmm3[13],xmm30[13],xmm3[14],xmm30[14],xmm3[15],xmm30[15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm18 = xmm18[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm18, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm18 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm31, %zmm6, %zmm18
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
; AVX512DQ-BW-FCP-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k2}
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm0 {%k3}
; AVX512DQ-BW-FCP-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm23 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm15, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm14, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3],xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm18, %xmm4, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm4, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm17, %xmm2
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm16, %xmm0
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm2, %xmm0, %xmm0
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm16[0],xmm17[0],xmm16[1],xmm17[1],xmm16[2],xmm17[2],xmm16[3],xmm17[3],xmm16[4],xmm17[4],xmm16[5],xmm17[5],xmm16[6],xmm17[6],xmm16[7],xmm17[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm0
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm29 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm13, %xmm19, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm17, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm0, %xmm6, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm17[0],xmm19[0],xmm17[1],xmm19[1],xmm17[2],xmm19[2],xmm17[3],xmm19[3],xmm17[4],xmm19[4],xmm17[5],xmm19[5],xmm17[6],xmm19[6],xmm17[7],xmm19[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm6, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm22, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm21, %xmm8
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm21[0],xmm22[0],xmm21[1],xmm22[1],xmm21[2],xmm22[2],xmm21[3],xmm22[3],xmm21[4],xmm22[4],xmm21[5],xmm22[5],xmm21[6],xmm22[6],xmm21[7],xmm22[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm8
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm8, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm6[0,1,0,1,4,5,4,5]
; AVX512DQ-BW-FCP-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm0 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm2
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm2, %xmm2
-; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm10, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm2, %zmm3, %zmm3
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm8, %zmm0 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm16, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm15, %xmm7
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm7
+; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm6, %zmm7, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm31, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm7 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm8, %zmm7, %zmm7
; AVX512DQ-BW-FCP-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm1 {%k2}
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm6 {%k3}
; AVX512DQ-BW-FCP-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm24, %zmm28, %zmm1
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm0 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm13[28],zero,ymm13[30,31,30,31],zero,ymm13[29],zero,ymm13[31,28,29]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm6, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[30],zero,ymm7[28],zero,zero,zero,zero,ymm7[31],zero,ymm7[29],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[2,3,2,3],zmm2[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm30[0],xmm3[1],xmm30[1],xmm3[2],xmm30[2],xmm3[3],xmm30[3],xmm3[4],xmm30[4],xmm3[5],xmm30[5],xmm3[6],xmm30[6],xmm3[7],xmm30[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero
+; AVX512DQ-BW-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29]
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[2,3,2,3],zmm1[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm31, %zmm3, %zmm3
+; AVX512DQ-BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm26, %zmm1
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,19,20,21],zero,zmm1[19],zero,zmm1[21,20,21,22],zero,zmm1[20],zero,zmm1[22,23,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero,zero,zero,zmm1[59],zero
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm27, %zmm3
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero,zmm3[59],zero,zmm3[57]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm3, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm24, %zmm3
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[18],zero,zmm3[18,19,20,21],zero,zmm3[19],zero,zmm3[25,26,27,22],zero,zmm3[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm3[56,57],zero,zmm3[55],zero,zmm3[53,54,55,58],zero,zmm3[56],zero,zmm3[60,61,58,59]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm4 # 32-byte Folded Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm23, %zmm4
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22],zero,zmm4[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero,zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm3, %zmm4, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm20, %zmm30, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm1
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm1[56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm27, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm25, %zmm4
; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm4[20],zero,zmm4[18],zero,zero,zero,zero,zmm4[21],zero,zmm4[19],zero,zero,zero,zero,zmm4[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm4[57],zero,zmm4[55],zero,zero,zero,zero,zmm4[58],zero,zmm4[56],zero,zero
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm4, %zmm1
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm2, %zmm4, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm8, %zmm4, %zmm4
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, 256(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 384(%rax)
-; AVX512DQ-BW-FCP-NEXT: addq $40, %rsp
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 256(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 384(%rax)
+; AVX512DQ-BW-FCP-NEXT: addq $104, %rsp
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
More information about the llvm-commits
mailing list