[llvm] [X86] combineConcatVectorOps - add concatenation handling for BITCAST nodes (PR #133913)
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 10 08:34:06 PDT 2025
https://github.com/RKSimon updated https://github.com/llvm/llvm-project/pull/133913
>From f184621f927c8a9f34d0eb577635601acf27393e Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Tue, 1 Apr 2025 14:22:36 +0100
Subject: [PATCH] [X86] combineConcatVectorOps - add concatenation handling for
BITCAST nodes
These nodes are effectively free. For bitcasts we should only concatenate if the inner nodes will concatenate together.
This also exposed a regression in canonicalizeShuffleWithOp that failed to realise it could potentially merge shuffles with a CONCAT_VECTORS node.
---
llvm/lib/Target/X86/X86ISelLowering.cpp | 25 +
llvm/test/CodeGen/X86/shift-i512.ll | 40 +-
llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll | 18 +-
llvm/test/CodeGen/X86/vector-fshr-256.ll | 48 +-
llvm/test/CodeGen/X86/vector-fshr-rot-256.ll | 25 +-
.../vector-interleaved-store-i16-stride-4.ll | 30 +-
.../vector-interleaved-store-i16-stride-7.ll | 64 +-
.../vector-interleaved-store-i16-stride-8.ll | 88 +-
.../vector-interleaved-store-i8-stride-5.ll | 416 +-
.../vector-interleaved-store-i8-stride-6.ll | 596 +-
.../vector-interleaved-store-i8-stride-7.ll | 5344 ++++++++---------
llvm/test/CodeGen/X86/vector-reduce-ctpop.ll | 79 +-
llvm/test/CodeGen/X86/widen_fdiv.ll | 44 +-
13 files changed, 3349 insertions(+), 3468 deletions(-)
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a3c423270f44a..42ef817e01456 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -41704,6 +41704,7 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
getTargetConstantFromNode(dyn_cast<LoadSDNode>(Op)) ||
(Op.getOpcode() == Opc && Op->hasOneUse()) ||
(Op.getOpcode() == ISD::INSERT_SUBVECTOR && Op->hasOneUse()) ||
+ (Op.getOpcode() == ISD::CONCAT_VECTORS && Op->hasOneUse()) ||
(FoldShuf && isTargetShuffle(Op.getOpcode()) && Op->hasOneUse()) ||
DAG.isSplatValue(Op, /*AllowUndefs*/ false);
};
@@ -58134,6 +58135,30 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
unsigned Opcode = Op0.getOpcode();
switch (Opcode) {
+ case ISD::BITCAST: {
+ // TODO: Support AVX1/AVX2 bitcasts.
+ SmallVector<SDValue, 4> SubOps;
+ for (SDValue SubOp : Ops)
+ SubOps.push_back(peekThroughBitcasts(SubOp.getOperand(0)));
+ EVT InnerVT = SubOps[0].getValueType();
+ unsigned InnerSizeInBits = InnerVT.getScalarSizeInBits();
+ if (!IsSplat && InnerVT.isSimple() && InnerVT.isVector() &&
+ (Subtarget.hasBWI() ||
+ (EltSizeInBits >= 32 && InnerSizeInBits >= 32)) &&
+ ((VT.is256BitVector() && Subtarget.hasVLX()) ||
+ (VT.is512BitVector() && Subtarget.useAVX512Regs())) &&
+ llvm::all_of(SubOps, [InnerVT](SDValue Op) {
+ return Op.getValueType() == InnerVT;
+ })) {
+ MVT ConcatSVT = InnerVT.getScalarType().getSimpleVT();
+ MVT ConcatVT = MVT::getVectorVT(
+ ConcatSVT, VT.getSizeInBits() / ConcatSVT.getSizeInBits());
+ if (SDValue ConcatSrc = combineConcatVectorOps(
+ DL, ConcatVT, SubOps, DAG, Subtarget, Depth + 1))
+ return DAG.getBitcast(VT, ConcatSrc);
+ }
+ break;
+ }
case ISD::VECTOR_SHUFFLE: {
// TODO: Generalize NumOps support.
if (!IsSplat && NumOps == 2 &&
diff --git a/llvm/test/CodeGen/X86/shift-i512.ll b/llvm/test/CodeGen/X86/shift-i512.ll
index f7dd1dc0949f5..c7da04171e6a1 100644
--- a/llvm/test/CodeGen/X86/shift-i512.ll
+++ b/llvm/test/CodeGen/X86/shift-i512.ll
@@ -123,20 +123,18 @@ define <8 x i64> @lshr_i512_1(<8 x i64> %a) {
;
; AVX512VBMI-LABEL: lshr_i512_1:
; AVX512VBMI: # %bb.0:
-; AVX512VBMI-NEXT: vextracti32x4 $3, %zmm0, %xmm1
-; AVX512VBMI-NEXT: vextracti32x4 $2, %zmm0, %xmm2
-; AVX512VBMI-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX512VBMI-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512VBMI-NEXT: vextracti32x4 $3, %zmm0, %xmm2
+; AVX512VBMI-NEXT: vextracti32x4 $2, %zmm0, %xmm3
; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[2,3,2,3]
; AVX512VBMI-NEXT: vpshldq $63, %xmm4, %xmm2, %xmm4
-; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
-; AVX512VBMI-NEXT: vpshldq $63, %xmm5, %xmm3, %xmm3
-; AVX512VBMI-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
-; AVX512VBMI-NEXT: vpshldq $63, %xmm2, %xmm1, %xmm2
-; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; AVX512VBMI-NEXT: vpsrlq $1, %xmm1, %xmm1
-; AVX512VBMI-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512VBMI-NEXT: vpsrlq $1, %xmm2, %xmm2
+; AVX512VBMI-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm2
+; AVX512VBMI-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX512VBMI-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[2,3,2,3,6,7,6,7]
+; AVX512VBMI-NEXT: vpshldq $63, %ymm3, %ymm1, %ymm1
+; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VBMI-NEXT: vpshufd {{.*#+}} zmm2 = zmm0[2,3,2,3,6,7,6,7,10,11,10,11,14,15,14,15]
; AVX512VBMI-NEXT: vpshldq $63, %zmm0, %zmm2, %zmm0
; AVX512VBMI-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
@@ -238,20 +236,18 @@ define <8 x i64> @ashr_i512_1(<8 x i64> %a) {
;
; AVX512VBMI-LABEL: ashr_i512_1:
; AVX512VBMI: # %bb.0:
-; AVX512VBMI-NEXT: vextracti32x4 $3, %zmm0, %xmm1
-; AVX512VBMI-NEXT: vextracti32x4 $2, %zmm0, %xmm2
-; AVX512VBMI-NEXT: vextracti128 $1, %ymm0, %xmm3
+; AVX512VBMI-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512VBMI-NEXT: vextracti32x4 $3, %zmm0, %xmm2
+; AVX512VBMI-NEXT: vextracti32x4 $2, %zmm0, %xmm3
; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[2,3,2,3]
; AVX512VBMI-NEXT: vpshldq $63, %xmm4, %xmm2, %xmm4
-; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
-; AVX512VBMI-NEXT: vpshldq $63, %xmm5, %xmm3, %xmm3
-; AVX512VBMI-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
-; AVX512VBMI-NEXT: vpshldq $63, %xmm2, %xmm1, %xmm2
-; AVX512VBMI-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; AVX512VBMI-NEXT: vpsraq $1, %xmm1, %xmm1
-; AVX512VBMI-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512VBMI-NEXT: vpsraq $1, %xmm2, %xmm2
+; AVX512VBMI-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm2
+; AVX512VBMI-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX512VBMI-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[2,3,2,3,6,7,6,7]
+; AVX512VBMI-NEXT: vpshldq $63, %ymm3, %ymm1, %ymm1
+; AVX512VBMI-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VBMI-NEXT: vpshufd {{.*#+}} zmm2 = zmm0[2,3,2,3,6,7,6,7,10,11,10,11,14,15,14,15]
; AVX512VBMI-NEXT: vpshldq $63, %zmm0, %zmm2, %zmm0
; AVX512VBMI-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
index ec442c185706c..e27a77ed2293d 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
@@ -297,23 +297,21 @@ define <16 x i8> @trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_
;
; AVX512BW-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
; AVX512BW: # %bb.0:
-; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm1 = [0,4,16,21]
-; AVX512BW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
-; AVX512BW-NEXT: vpermt2d %zmm2, %zmm1, %zmm0
+; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm1 = [0,4,8,13]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u,33,37,41,45,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,49,53,57,62,u,u,u,u,u,u,u,u]
+; AVX512BW-NEXT: vpermd %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BWVL-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
; AVX512BWVL: # %bb.0:
-; AVX512BWVL-NEXT: vpmovsxbd {{.*#+}} xmm1 = [0,4,8,13]
+; AVX512BWVL-NEXT: vpmovsxbd {{.*#+}} xmm1 = [8,12,0,5]
; AVX512BWVL-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512BWVL-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
-; AVX512BWVL-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
-; AVX512BWVL-NEXT: vpermt2d %ymm2, %ymm1, %ymm0
-; AVX512BWVL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX512BWVL-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512BWVL-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u,33,37,41,45,u,u,u,u,u,u,u,u,u,u,u,u,49,53,57,61,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512BWVL-NEXT: vpermd %zmm0, %zmm1, %zmm0
+; AVX512BWVL-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; AVX512BWVL-NEXT: vzeroupper
; AVX512BWVL-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-fshr-256.ll b/llvm/test/CodeGen/X86/vector-fshr-256.ll
index 35c707eac83b4..0fa2c858ff000 100644
--- a/llvm/test/CodeGen/X86/vector-fshr-256.ll
+++ b/llvm/test/CodeGen/X86/vector-fshr-256.ll
@@ -1228,13 +1228,14 @@ define <32 x i8> @splatvar_funnnel_v32i8(<32 x i8> %x, <32 x i8> %y, <32 x i8> %
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v32i8:
; AVX512VBMI2: # %bb.0:
-; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78,16,18,20,22,24,26,28,30,80,82,84,86,88,90,92,94]
-; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15],ymm1[24],ymm0[24],ymm1[25],ymm0[25],ymm1[26],ymm0[26],ymm1[27],ymm0[27],ymm1[28],ymm0[28],ymm1[29],ymm0[29],ymm1[30],ymm0[30],ymm1[31],ymm0[31]
-; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
-; AVX512VBMI2-NEXT: vpsrlw %xmm2, %ymm4, %ymm4
-; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
-; AVX512VBMI2-NEXT: vpsrlw %xmm2, %ymm0, %ymm0
-; AVX512VBMI2-NEXT: vpermt2b %zmm4, %zmm3, %zmm0
+; AVX512VBMI2-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
+; AVX512VBMI2-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,64,1,65,2,66,3,67,4,68,5,69,6,70,7,71,16,80,17,81,18,82,19,83,20,84,21,85,22,86,23,87,8,72,9,73,10,74,11,75,12,76,13,77,14,78,15,79,24,88,25,89,26,90,27,91,28,92,29,93,30,94,31,95]
+; AVX512VBMI2-NEXT: vpermi2b %zmm0, %zmm1, %zmm3
+; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm0
+; AVX512VBMI2-NEXT: vpsrlw %xmm0, %zmm3, %zmm0
+; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} ymm1 = [0,2,4,6,8,10,12,14,32,34,36,38,40,42,44,46,16,18,20,22,24,26,28,30,48,50,52,54,56,58,60,62]
+; AVX512VBMI2-NEXT: vpermb %zmm0, %zmm1, %zmm0
; AVX512VBMI2-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
; AVX512VBMI2-NEXT: retq
;
@@ -1251,16 +1252,29 @@ define <32 x i8> @splatvar_funnnel_v32i8(<32 x i8> %x, <32 x i8> %y, <32 x i8> %
; AVX512VLBW-NEXT: vpackuswb %ymm3, %ymm0, %ymm0
; AVX512VLBW-NEXT: retq
;
-; AVX10-LABEL: splatvar_funnnel_v32i8:
-; AVX10: # %bb.0:
-; AVX10-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15],ymm1[24],ymm0[24],ymm1[25],ymm0[25],ymm1[26],ymm0[26],ymm1[27],ymm0[27],ymm1[28],ymm0[28],ymm1[29],ymm0[29],ymm1[30],ymm0[30],ymm1[31],ymm0[31]
-; AVX10-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
-; AVX10-NEXT: vpsrlw %xmm2, %ymm3, %ymm3
-; AVX10-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
-; AVX10-NEXT: vpsrlw %xmm2, %ymm0, %ymm1
-; AVX10-NEXT: vmovdqa {{.*#+}} ymm0 = [0,2,4,6,8,10,12,14,32,34,36,38,40,42,44,46,16,18,20,22,24,26,28,30,48,50,52,54,56,58,60,62]
-; AVX10-NEXT: vpermi2b %ymm3, %ymm1, %ymm0
-; AVX10-NEXT: retq
+; AVX512VLVBMI2-LABEL: splatvar_funnnel_v32i8:
+; AVX512VLVBMI2: # %bb.0:
+; AVX512VLVBMI2-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
+; AVX512VLVBMI2-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [8,72,9,73,10,74,11,75,12,76,13,77,14,78,15,79,24,88,25,89,26,90,27,91,28,92,29,93,30,94,31,95,0,64,1,65,2,66,3,67,4,68,5,69,6,70,7,71,16,80,17,81,18,82,19,83,20,84,21,85,22,86,23,87]
+; AVX512VLVBMI2-NEXT: vpermi2b %zmm0, %zmm1, %zmm3
+; AVX512VLVBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm0
+; AVX512VLVBMI2-NEXT: vpsrlw %xmm0, %zmm3, %zmm0
+; AVX512VLVBMI2-NEXT: vmovdqa {{.*#+}} ymm1 = [32,34,36,38,40,42,44,46,0,2,4,6,8,10,12,14,48,50,52,54,56,58,60,62,16,18,20,22,24,26,28,30]
+; AVX512VLVBMI2-NEXT: vpermb %zmm0, %zmm1, %zmm0
+; AVX512VLVBMI2-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
+; AVX512VLVBMI2-NEXT: retq
+;
+; AVX10_256-LABEL: splatvar_funnnel_v32i8:
+; AVX10_256: # %bb.0:
+; AVX10_256-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15],ymm1[24],ymm0[24],ymm1[25],ymm0[25],ymm1[26],ymm0[26],ymm1[27],ymm0[27],ymm1[28],ymm0[28],ymm1[29],ymm0[29],ymm1[30],ymm0[30],ymm1[31],ymm0[31]
+; AVX10_256-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
+; AVX10_256-NEXT: vpsrlw %xmm2, %ymm3, %ymm3
+; AVX10_256-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
+; AVX10_256-NEXT: vpsrlw %xmm2, %ymm0, %ymm1
+; AVX10_256-NEXT: vmovdqa {{.*#+}} ymm0 = [0,2,4,6,8,10,12,14,32,34,36,38,40,42,44,46,16,18,20,22,24,26,28,30,48,50,52,54,56,58,60,62]
+; AVX10_256-NEXT: vpermi2b %ymm3, %ymm1, %ymm0
+; AVX10_256-NEXT: retq
;
; XOPAVX1-LABEL: splatvar_funnnel_v32i8:
; XOPAVX1: # %bb.0:
diff --git a/llvm/test/CodeGen/X86/vector-fshr-rot-256.ll b/llvm/test/CodeGen/X86/vector-fshr-rot-256.ll
index 5cac190eae690..3d4f283260aa5 100644
--- a/llvm/test/CodeGen/X86/vector-fshr-rot-256.ll
+++ b/llvm/test/CodeGen/X86/vector-fshr-rot-256.ll
@@ -992,25 +992,26 @@ define <32 x i8> @splatvar_funnnel_v32i8(<32 x i8> %x, <32 x i8> %amt) nounwind
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v32i8:
; AVX512VBMI2: # %bb.0:
-; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78,16,18,20,22,24,26,28,30,80,82,84,86,88,90,92,94]
-; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512VBMI2-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14,32,34,36,38,40,42,44,46,16,18,20,22,24,26,28,30,48,50,52,54,56,58,60,62]
+; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512VBMI2-NEXT: vpermb %zmm0, %zmm3, %zmm0
; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
-; AVX512VBMI2-NEXT: vpsrlw %xmm1, %ymm3, %ymm3
-; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX512VBMI2-NEXT: vpsrlw %xmm1, %ymm0, %ymm0
-; AVX512VBMI2-NEXT: vpermt2b %zmm3, %zmm2, %zmm0
+; AVX512VBMI2-NEXT: vpsrlw %xmm1, %zmm0, %zmm0
+; AVX512VBMI2-NEXT: vpermb %zmm0, %zmm2, %zmm0
; AVX512VBMI2-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatvar_funnnel_v32i8:
; AVX512VLVBMI2: # %bb.0:
-; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512VLVBMI2-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512VLVBMI2-NEXT: vmovdqa {{.*#+}} ymm2 = [32,34,36,38,40,42,44,46,0,2,4,6,8,10,12,14,48,50,52,54,56,58,60,62,16,18,20,22,24,26,28,30]
+; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512VLVBMI2-NEXT: vpermb %zmm0, %zmm3, %zmm0
; AVX512VLVBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
-; AVX512VLVBMI2-NEXT: vpsrlw %xmm1, %ymm2, %ymm2
-; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX512VLVBMI2-NEXT: vpsrlw %xmm1, %ymm0, %ymm1
-; AVX512VLVBMI2-NEXT: vmovdqa {{.*#+}} ymm0 = [0,2,4,6,8,10,12,14,32,34,36,38,40,42,44,46,16,18,20,22,24,26,28,30,48,50,52,54,56,58,60,62]
-; AVX512VLVBMI2-NEXT: vpermi2b %ymm2, %ymm1, %ymm0
+; AVX512VLVBMI2-NEXT: vpsrlw %xmm1, %zmm0, %zmm0
+; AVX512VLVBMI2-NEXT: vpermb %zmm0, %zmm2, %zmm0
+; AVX512VLVBMI2-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
; AVX512VLVBMI2-NEXT: retq
;
; XOPAVX1-LABEL: splatvar_funnnel_v32i8:
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
index fc4377a08d560..3e04c2c8120cc 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
@@ -130,13 +130,14 @@ define void @store_i16_stride4_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512BW-FCP-LABEL: store_i16_stride4_vf2:
; AVX512BW-FCP: # %bb.0:
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [0,2,8,10,1,3,9,11]
-; AVX512BW-FCP-NEXT: vpermi2w %xmm1, %xmm0, %xmm2
-; AVX512BW-FCP-NEXT: vmovdqa %xmm2, (%r8)
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [8,24,0,16,9,25,1,17]
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%rdi), %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: vpermt2w %ymm1, %ymm2, %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa %xmm0, (%r8)
+; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
; AVX512DQ-BW-LABEL: store_i16_stride4_vf2:
@@ -152,13 +153,14 @@ define void @store_i16_stride4_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQ-BW-FCP-LABEL: store_i16_stride4_vf2:
; AVX512DQ-BW-FCP: # %bb.0:
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [0,2,8,10,1,3,9,11]
-; AVX512DQ-BW-FCP-NEXT: vpermi2w %xmm1, %xmm0, %xmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm2, (%r8)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [8,24,0,16,9,25,1,17]
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rdi), %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm1, %ymm2, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, (%r8)
+; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <2 x i16>, ptr %in.vecptr0, align 64
%in.vec1 = load <2 x i16>, ptr %in.vecptr1, align 64
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index 7b619344e83f6..3767f1d68dfb7 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
@@ -259,14 +259,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
+; AVX512BW-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512BW-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512BW-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
+; AVX512BW-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm1
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
; AVX512BW-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
; AVX512BW-NEXT: vextracti128 $1, %ymm2, %xmm0
; AVX512BW-NEXT: vpextrd $2, %xmm0, 24(%rax)
@@ -280,14 +280,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm1
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
; AVX512BW-FCP-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
; AVX512BW-FCP-NEXT: vextracti128 $1, %ymm2, %xmm0
; AVX512BW-FCP-NEXT: vpextrd $2, %xmm0, 24(%rax)
@@ -301,14 +301,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm1
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
; AVX512DQ-BW-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
; AVX512DQ-BW-NEXT: vextracti128 $1, %ymm2, %xmm0
; AVX512DQ-BW-NEXT: vpextrd $2, %xmm0, 24(%rax)
@@ -322,14 +322,14 @@ define void @store_i16_stride7_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,1,3,5,7,17,19,21,0,0]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,16,18,8,10,24,1,3,17,19,9,11,25,0,0]
; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
; AVX512DQ-BW-FCP-NEXT: vextracti128 $1, %ymm2, %xmm0
; AVX512DQ-BW-FCP-NEXT: vpextrd $2, %xmm0, 24(%rax)
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
index 0eaf71552bcbc..26b791c4e26e3 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll
@@ -231,17 +231,17 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512BW-NEXT: vmovdqa (%r11), %xmm3
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
-; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512BW-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512BW-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512BW-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512BW-NEXT: vinserti128 $1, (%r11), %ymm3, %ymm3
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512BW-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512BW-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm2
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,48,0,32,24,56,8,40,17,49,1,33,25,57,9,41]
+; AVX512BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT: vmovdqa %ymm2, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
@@ -252,17 +252,17 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm3
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
-; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512BW-FCP-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r11), %ymm3, %ymm3
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm2
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,48,0,32,24,56,8,40,17,49,1,33,25,57,9,41]
+; AVX512BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm2
; AVX512BW-FCP-NEXT: vmovdqa %ymm2, (%rax)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
@@ -273,17 +273,17 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm3
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
-; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-BW-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r11), %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512DQ-BW-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm2
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,48,0,32,24,56,8,40,17,49,1,33,25,57,9,41]
+; AVX512DQ-BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm2
; AVX512DQ-BW-NEXT: vmovdqa %ymm2, (%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
@@ -294,17 +294,17 @@ define void @store_i16_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]
-; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm1, %ymm0, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r11), %ymm3, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r8), %ymm0, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm2, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [16,48,0,32,24,56,8,40,17,49,1,33,25,57,9,41]
+; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm2, (%rax)
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-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 302da6ef63796..a9e25ef7caad8 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
@@ -4668,23 +4668,23 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; 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, %ymm20
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm8
; 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: vpshufb %ymm2, %ymm8, %ymm1
; AVX512-NEXT: vmovdqa64 %ymm2, %ymm21
; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm17
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm10
; 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: vpshufb %xmm1, %xmm10, %xmm0
; AVX512-NEXT: vmovdqa64 %xmm1, %xmm28
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm9
; 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: vpshufb %xmm2, %xmm9, %xmm1
; AVX512-NEXT: vmovdqa64 %xmm2, %xmm29
-; AVX512-NEXT: vporq %xmm0, %xmm1, %xmm19
+; AVX512-NEXT: vporq %xmm0, %xmm1, %xmm18
; 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 {{.*#+}} ymm5 = [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 %ymm5, %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
@@ -4698,16 +4698,16 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vpshufb %xmm2, %xmm7, %xmm1
; AVX512-NEXT: vmovdqa64 %xmm2, %xmm31
; AVX512-NEXT: vporq %xmm0, %xmm1, %xmm23
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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: # ymm14 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm14, %ymm5, %ymm0
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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: # ymm12 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm12, %ymm8, %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: vpshufb %ymm4, %ymm8, %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 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm8, %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
@@ -4717,121 +4717,123 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; 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, %ymm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm25
+; 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, %ymm9
+; 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, %ymm9, %zmm13, %zmm27
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm27
; AVX512-NEXT: vmovdqa (%rcx), %ymm13
-; AVX512-NEXT: vpshufb %ymm8, %ymm13, %ymm8
-; AVX512-NEXT: vmovdqa (%rdx), %ymm9
-; AVX512-NEXT: vpshufb %ymm3, %ymm9, %ymm3
-; AVX512-NEXT: vporq %ymm8, %ymm3, %ymm16
+; AVX512-NEXT: vpshufb %ymm5, %ymm13, %ymm5
+; AVX512-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512-NEXT: vporq %ymm5, %ymm3, %ymm16
; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512-NEXT: vpshufb %ymm15, %ymm9, %ymm3
-; AVX512-NEXT: vporq %ymm0, %ymm3, %ymm18
+; AVX512-NEXT: vpshufb %ymm15, %ymm14, %ymm3
+; AVX512-NEXT: vporq %ymm0, %ymm3, %ymm19
; AVX512-NEXT: vmovdqa (%rsi), %ymm3
; AVX512-NEXT: vmovdqa64 %ymm20, %ymm0
; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512-NEXT: vmovdqa (%rdi), %ymm8
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm15
-; AVX512-NEXT: vpshufb %ymm15, %ymm8, %ymm15
-; AVX512-NEXT: vpor %ymm0, %ymm15, %ymm15
-; AVX512-NEXT: vpshufb %ymm4, %ymm8, %ymm0
+; AVX512-NEXT: vmovdqa (%rdi), %ymm15
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm5
+; AVX512-NEXT: vpshufb %ymm5, %ymm15, %ymm5
+; AVX512-NEXT: vpor %ymm0, %ymm5, %ymm5
+; AVX512-NEXT: vpshufb %ymm4, %ymm15, %ymm0
; AVX512-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512-NEXT: vpshufb %ymm11, %ymm9, %ymm2
+; AVX512-NEXT: vpor %ymm0, %ymm2, %ymm2
+; AVX512-NEXT: vpshufb %ymm11, %ymm14, %ymm0
; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512-NEXT: vporq %ymm2, %ymm1, %ymm20
-; AVX512-NEXT: vpshufb %ymm14, %ymm8, %ymm1
-; AVX512-NEXT: vpshufb %ymm5, %ymm3, %ymm2
-; AVX512-NEXT: vporq %ymm1, %ymm2, %ymm21
-; AVX512-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm20
+; AVX512-NEXT: vpshufb %ymm12, %ymm15, %ymm0
+; AVX512-NEXT: vpshufb %ymm8, %ymm3, %ymm1
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512-NEXT: vmovdqa (%rdi), %xmm0
; AVX512-NEXT: vmovdqa (%rsi), %xmm8
; AVX512-NEXT: vmovdqa64 %xmm28, %xmm1
-; AVX512-NEXT: vpshufb %xmm1, %xmm5, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm29, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm2
-; AVX512-NEXT: vpor %xmm1, %xmm2, %xmm3
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512-NEXT: vpshufb %xmm3, %xmm8, %xmm3
+; AVX512-NEXT: vpor %xmm1, %xmm3, %xmm11
; AVX512-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512-NEXT: vmovdqa (%rdx), %xmm3
; AVX512-NEXT: vmovdqa64 %xmm30, %xmm4
; AVX512-NEXT: vpshufb %xmm4, %xmm1, %xmm4
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm9
-; AVX512-NEXT: vpshufb %xmm9, %xmm2, %xmm9
-; AVX512-NEXT: vpor %xmm4, %xmm9, %xmm9
+; AVX512-NEXT: vmovdqa64 %xmm31, %xmm12
+; AVX512-NEXT: vpshufb %xmm12, %xmm3, %xmm12
+; AVX512-NEXT: vpor %xmm4, %xmm12, %xmm12
; AVX512-NEXT: vmovdqa 32(%r8), %ymm4
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm4, %ymm4
-; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = mem[1,1,2,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,1,1]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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: vpandn %ymm13, %ymm14, %ymm13
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm13, %zmm4
-; AVX512-NEXT: vmovdqa (%r8), %ymm13
-; AVX512-NEXT: vpshufb %ymm11, %ymm13, %ymm11
-; AVX512-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,1,1,4,6,5,5]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,3,2]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
-; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm11
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm4, %ymm4
+; AVX512-NEXT: vpshufd {{.*#+}} xmm14 = mem[1,1,2,2]
+; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,1,1]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm15 = [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: vpandn %ymm14, %ymm15, %ymm14
+; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm14, %zmm4
+; AVX512-NEXT: vmovdqa (%r8), %ymm14
+; AVX512-NEXT: vpshufb %ymm13, %ymm14, %ymm13
+; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,2,1,1,4,6,5,5]
+; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,3,2]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm14, %ymm14
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
; 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: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3],xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
-; AVX512-NEXT: vpshufb %xmm12, %xmm5, %xmm5
-; AVX512-NEXT: vinserti32x4 $2, %xmm3, %zmm5, %zmm3
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-NEXT: vpshufb %xmm10, %xmm9, %xmm9
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
+; AVX512-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512-NEXT: vinserti32x4 $2, %xmm11, %zmm0, %zmm0
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512-NEXT: vinserti32x4 $2, %xmm9, %zmm1, %zmm1
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm19[0,0,1,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm2, %zmm2
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm23[0,0,1,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm22, %zmm5, %zmm5
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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 {{.*#+}} zmm5 = zmm2 ^ (zmm7 & (zmm5 ^ zmm2))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm5 & zmm14)
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm18[2,2,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm16, %zmm2
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm7 & (zmm0 ^ zmm2))
-; AVX512-NEXT: vporq %zmm24, %zmm26, %zmm2
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vporq %zmm25, %zmm27, %zmm5
-; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vinserti32x4 $2, %xmm12, %zmm1, %zmm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; 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 {{.*#+}} zmm7 = zmm3 ^ (zmm8 & (zmm7 ^ zmm3))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm7 & zmm15)
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm19[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm16, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm3 ^ (zmm8 & (zmm2 ^ zmm3))
+; AVX512-NEXT: vpermq {{.*#+}} zmm3 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vporq %zmm3, %zmm5, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vporq %zmm5, %zmm7, %zmm5
; AVX512-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-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm7 & (zmm5 ^ zmm2))
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm20[2,2,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm7 & (zmm5 ^ zmm3))
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,2,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm21[2,2,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm2 ^ (zmm7 & (zmm6 ^ zmm2))
-; AVX512-NEXT: vmovdqa64 (%r8), %zmm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm6
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm3 ^ (zmm7 & (zmm6 ^ zmm3))
+; AVX512-NEXT: vmovdqa64 (%r8), %zmm3
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm7 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
-; AVX512-NEXT: vpermd %zmm2, %zmm7, %zmm7
+; AVX512-NEXT: vpermd %zmm3, %zmm7, %zmm7
; AVX512-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 | (zmm0 & mem)
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512-NEXT: vpermd %zmm2, %zmm0, %zmm0
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
-; AVX512-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 | (zmm2 & mem)
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512-NEXT: vpermd %zmm3, %zmm2, %zmm2
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm6))
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3))
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512-NEXT: vpermd %zmm2, %zmm3, %zmm2
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512-NEXT: vmovdqa64 %zmm11, 64(%r9)
-; AVX512-NEXT: vmovdqa64 %zmm2, (%r9)
-; AVX512-NEXT: vmovdqa64 %zmm0, 128(%r9)
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512-NEXT: vpermd %zmm3, %zmm0, %zmm0
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
+; AVX512-NEXT: vmovdqa64 %zmm13, 64(%r9)
+; AVX512-NEXT: vmovdqa64 %zmm0, (%r9)
+; AVX512-NEXT: vmovdqa64 %zmm2, 128(%r9)
; AVX512-NEXT: vmovdqa64 %zmm7, 256(%r9)
; AVX512-NEXT: vmovdqa64 %zmm4, 192(%r9)
; AVX512-NEXT: vzeroupper
@@ -5005,23 +5007,23 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; 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, %ymm20
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm8
; 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: vpshufb %ymm2, %ymm8, %ymm1
; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm21
; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm17
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm10
; 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: vpshufb %xmm1, %xmm10, %xmm0
; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm28
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm9
; 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: vpshufb %xmm2, %xmm9, %xmm1
; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm29
-; AVX512DQ-NEXT: vporq %xmm0, %xmm1, %xmm19
+; AVX512DQ-NEXT: vporq %xmm0, %xmm1, %xmm18
; 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 {{.*#+}} ymm5 = [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 %ymm5, %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
@@ -5035,16 +5037,16 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vpshufb %xmm2, %xmm7, %xmm1
; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm31
; AVX512DQ-NEXT: vporq %xmm0, %xmm1, %xmm23
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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: # ymm14 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm5, %ymm0
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm8, %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: vpshufb %ymm4, %ymm8, %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 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm8, %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
@@ -5054,121 +5056,123 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; 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, %ymm9
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm25
+; 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, %ymm9
+; 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, %ymm9, %zmm13, %zmm27
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm27
; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm8, %ymm13, %ymm8
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm9
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm9, %ymm3
-; AVX512DQ-NEXT: vporq %ymm8, %ymm3, %ymm16
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm13, %ymm5
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512DQ-NEXT: vporq %ymm5, %ymm3, %ymm16
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm9, %ymm3
-; AVX512DQ-NEXT: vporq %ymm0, %ymm3, %ymm18
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm14, %ymm3
+; AVX512DQ-NEXT: vporq %ymm0, %ymm3, %ymm19
; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm3
; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm0
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm8
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm15
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm8, %ymm15
-; AVX512DQ-NEXT: vpor %ymm0, %ymm15, %ymm15
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm8, %ymm0
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm15
+; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm15, %ymm5
+; AVX512DQ-NEXT: vpor %ymm0, %ymm5, %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm15, %ymm0
; AVX512DQ-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512DQ-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm9, %ymm2
+; AVX512DQ-NEXT: vpor %ymm0, %ymm2, %ymm2
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm14, %ymm0
; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512DQ-NEXT: vporq %ymm2, %ymm1, %ymm20
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm8, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm5, %ymm3, %ymm2
-; AVX512DQ-NEXT: vporq %ymm1, %ymm2, %ymm21
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm20
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm15, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm3, %ymm1
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0
; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm8
; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm5, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm8, %xmm2
-; AVX512DQ-NEXT: vpor %xmm1, %xmm2, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm8, %xmm3
+; AVX512DQ-NEXT: vpor %xmm1, %xmm3, %xmm11
; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm3
; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm4
; AVX512DQ-NEXT: vpshufb %xmm4, %xmm1, %xmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm9
-; AVX512DQ-NEXT: vpshufb %xmm9, %xmm2, %xmm9
-; AVX512DQ-NEXT: vpor %xmm4, %xmm9, %xmm9
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm12
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm3, %xmm12
+; AVX512DQ-NEXT: vpor %xmm4, %xmm12, %xmm12
; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm4
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm4, %ymm4
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm13 = mem[1,1,2,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,1,1]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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: vpandn %ymm13, %ymm14, %ymm13
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm13, %zmm4
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm13, %ymm11
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,1,1,4,6,5,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,3,2]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm11
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm4, %ymm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm14 = mem[1,1,2,2]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,1,1]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm15 = [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: vpandn %ymm14, %ymm15, %ymm14
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm14, %zmm4
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm14
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm14, %ymm13
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,2,1,1,4,6,5,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,3,2]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm14, %ymm14
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
; 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: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3],xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm5, %xmm5
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm3, %zmm5, %zmm3
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-NEXT: vpshufb %xmm10, %xmm9, %xmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
+; AVX512DQ-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm11, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm9, %zmm1, %zmm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm19[0,0,1,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm2, %zmm2
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm23[0,0,1,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm22, %zmm5, %zmm5
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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 {{.*#+}} zmm5 = zmm2 ^ (zmm7 & (zmm5 ^ zmm2))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm5 & zmm14)
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm18[2,2,3,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm16, %zmm2
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm7 & (zmm0 ^ zmm2))
-; AVX512DQ-NEXT: vporq %zmm24, %zmm26, %zmm2
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vporq %zmm25, %zmm27, %zmm5
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm12, %zmm1, %zmm1
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; 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 {{.*#+}} zmm7 = zmm3 ^ (zmm8 & (zmm7 ^ zmm3))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm7 & zmm15)
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm19[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm16, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm3 ^ (zmm8 & (zmm2 ^ zmm3))
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm3 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vporq %zmm3, %zmm5, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vporq %zmm5, %zmm7, %zmm5
; AVX512DQ-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-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm7 & (zmm5 ^ zmm2))
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm20[2,2,3,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm7 & (zmm5 ^ zmm3))
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,2,3,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm21[2,2,3,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm2 ^ (zmm7 & (zmm6 ^ zmm2))
-; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm6
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm3 ^ (zmm7 & (zmm6 ^ zmm3))
+; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm3
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm7 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
-; AVX512DQ-NEXT: vpermd %zmm2, %zmm7, %zmm7
+; AVX512DQ-NEXT: vpermd %zmm3, %zmm7, %zmm7
; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 | (zmm0 & mem)
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512DQ-NEXT: vpermd %zmm2, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm6))
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 | (zmm2 & mem)
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512DQ-NEXT: vpermd %zmm3, %zmm2, %zmm2
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm6))
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,1,4,4,5,5]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3))
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512DQ-NEXT: vpermd %zmm2, %zmm3, %zmm2
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512DQ-NEXT: vmovdqa64 %zmm11, 64(%r9)
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, (%r9)
-; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%r9)
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512DQ-NEXT: vpermd %zmm3, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
+; AVX512DQ-NEXT: vmovdqa64 %zmm13, 64(%r9)
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%r9)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 128(%r9)
; AVX512DQ-NEXT: vmovdqa64 %zmm7, 256(%r9)
; AVX512DQ-NEXT: vmovdqa64 %zmm4, 192(%r9)
; AVX512DQ-NEXT: vzeroupper
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
index 5c30c4cdf94d3..86e249649a291 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
@@ -6388,7 +6388,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: kmovd %r10d, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm9, %zmm0 {%k1}
; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm7[4,5,6,7,4,5,6,7]
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm9[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm14 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15]
+; AVX512BW-NEXT: vpshufb %zmm14, %zmm9, %zmm9
; AVX512BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
; AVX512BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512BW-NEXT: kmovd %r10d, %k2
@@ -6420,7 +6421,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: vmovdqu16 %zmm8, %zmm5 {%k1}
; AVX512BW-NEXT: vmovdqa (%r8), %ymm13
; AVX512BW-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm7 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512BW-NEXT: vpshufb %zmm14, %zmm7, %zmm7
; AVX512BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,2,2,3,6,6,6,7]
; AVX512BW-NEXT: vmovdqu16 %zmm7, %zmm5 {%k2}
; AVX512BW-NEXT: vmovdqa (%r9), %ymm12
@@ -6545,184 +6546,185 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512BW-FCP-LABEL: store_i8_stride6_vf64:
; AVX512BW-FCP: # %bb.0:
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm9
-; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm10
-; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm2
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[16],ymm1[16],ymm2[17],ymm1[17],ymm2[18],ymm1[18],ymm2[19],ymm1[19],ymm2[20],ymm1[20],ymm2[21],ymm1[21],ymm2[22],ymm1[22],ymm2[23],ymm1[23]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm1[2,2,2,3]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15],ymm10[24],ymm9[24],ymm10[25],ymm9[25],ymm10[26],ymm9[26],ymm10[27],ymm9[27],ymm10[28],ymm9[28],ymm10[29],ymm9[29],ymm10[30],ymm9[30],ymm10[31],ymm9[31]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-FCP-NEXT: vpermw %ymm3, %ymm0, %ymm3
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm4
-; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm18
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm21
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm21, %xmm8
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm30
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm24
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm24, %xmm11
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm24[0],xmm21[0],xmm24[1],xmm21[1],xmm24[2],xmm21[2],xmm24[3],xmm21[3],xmm24[4],xmm21[4],xmm24[5],xmm21[5],xmm24[6],xmm21[6],xmm24[7],xmm21[7]
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm1 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm7
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm2
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[16],ymm0[16],ymm2[17],ymm0[17],ymm2[18],ymm0[18],ymm2[19],ymm0[19],ymm2[20],ymm0[20],ymm2[21],ymm0[21],ymm2[22],ymm0[22],ymm2[23],ymm0[23]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-FCP-NEXT: vpermw %ymm3, %ymm1, %ymm3
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm5
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
+; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm4
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[16],ymm0[16],ymm4[17],ymm0[17],ymm4[18],ymm0[18],ymm4[19],ymm0[19],ymm4[20],ymm0[20],ymm4[21],ymm0[21],ymm4[22],ymm0[22],ymm4[23],ymm0[23]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15],ymm11[24],ymm10[24],ymm11[25],ymm10[25],ymm11[26],ymm10[26],ymm11[27],ymm10[27],ymm11[28],ymm10[28],ymm11[29],ymm10[29],ymm11[30],ymm10[30],ymm11[31],ymm10[31]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512BW-FCP-NEXT: vpermw %ymm8, %ymm4, %ymm8
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT: movl $613566756, %eax # imm = 0x24924924
+; AVX512BW-FCP-NEXT: kmovd %eax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm0 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm5
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm23
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm17
+; AVX512BW-FCP-NEXT: vpshufb %xmm3, %xmm17, %xmm8
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm25
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm21
+; AVX512BW-FCP-NEXT: vpshufb %xmm3, %xmm21, %xmm9
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm21[0],xmm17[0],xmm21[1],xmm17[1],xmm21[2],xmm17[2],xmm21[3],xmm17[3],xmm21[4],xmm17[4],xmm21[5],xmm17[5],xmm21[6],xmm17[6],xmm21[7],xmm17[7]
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm13 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm8, %zmm13, %zmm11
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm25
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19
+; AVX512BW-FCP-NEXT: vpermt2w %zmm8, %zmm13, %zmm9
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm30
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm18
; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm15 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm19, %xmm8
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm26
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm22
-; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm22, %xmm12
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm18, %xmm8
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm31
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm24
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm24, %xmm12
; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm22[0],xmm19[0],xmm22[1],xmm19[1],xmm22[2],xmm19[2],xmm22[3],xmm19[3],xmm22[4],xmm19[4],xmm22[5],xmm19[5],xmm22[6],xmm19[6],xmm22[7],xmm19[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm24[0],xmm18[0],xmm24[1],xmm18[1],xmm24[2],xmm18[2],xmm24[3],xmm18[3],xmm24[4],xmm18[4],xmm24[5],xmm18[5],xmm24[6],xmm18[6],xmm24[7],xmm18[7]
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
; AVX512BW-FCP-NEXT: vpermt2w %zmm12, %zmm16, %zmm8
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %xmm29
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %xmm26
; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm20
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm17 = [8,9,0,0,0,5,6,7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm20, %xmm14
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm19 = [8,9,0,0,0,5,6,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm20, %xmm14
; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm12 = xmm20[0],zero,xmm20[1],zero,xmm20[2],zero,xmm20[3],zero,xmm20[4],zero,xmm20[5],zero,xmm20[6],zero,xmm20[7],zero
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm27 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm14, %zmm27, %zmm12
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %xmm28
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm23
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm18, %xmm14
-; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm30, %xmm31
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm31 = xmm31[0],xmm14[0],xmm31[1],xmm14[1],xmm31[2],xmm14[2],xmm31[3],xmm14[3],xmm31[4],xmm14[4],xmm31[5],xmm14[5],xmm31[6],xmm14[6],xmm31[7],xmm14[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm30[0],xmm18[0],xmm30[1],xmm18[1],xmm30[2],xmm18[2],xmm30[3],xmm18[3],xmm30[4],xmm18[4],xmm30[5],xmm18[5],xmm30[6],xmm18[6],xmm30[7],xmm18[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm31, %zmm13, %zmm14
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm31 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm25, %xmm13
-; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm26, %xmm15
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm14, %zmm28, %zmm12
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %xmm27
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512BW-FCP-NEXT: vpshufb %xmm3, %xmm23, %xmm14
+; AVX512BW-FCP-NEXT: vpshufb %xmm3, %xmm25, %xmm29
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm29 = xmm29[0],xmm14[0],xmm29[1],xmm14[1],xmm29[2],xmm14[2],xmm29[3],xmm14[3],xmm29[4],xmm14[4],xmm29[5],xmm14[5],xmm29[6],xmm14[6],xmm29[7],xmm14[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm25[0],xmm23[0],xmm25[1],xmm23[1],xmm25[2],xmm23[2],xmm25[3],xmm23[3],xmm25[4],xmm23[4],xmm25[5],xmm23[5],xmm25[6],xmm23[6],xmm25[7],xmm23[7]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm29, %zmm13, %zmm14
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm29 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm30, %xmm13
+; AVX512BW-FCP-NEXT: vpshufb %xmm15, %xmm31, %xmm15
; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm26[0],xmm25[0],xmm26[1],xmm25[1],xmm26[2],xmm25[2],xmm26[3],xmm25[3],xmm26[4],xmm25[4],xmm26[5],xmm25[5],xmm26[6],xmm25[6],xmm26[7],xmm25[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm31[0],xmm30[0],xmm31[1],xmm30[1],xmm31[2],xmm30[2],xmm31[3],xmm30[3],xmm31[4],xmm30[4],xmm31[5],xmm30[5],xmm31[6],xmm30[6],xmm31[7],xmm30[7]
; AVX512BW-FCP-NEXT: vpermt2w %zmm15, %zmm16, %zmm13
-; AVX512BW-FCP-NEXT: vpshufb %xmm31, %xmm23, %xmm16
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm23[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm16, %zmm27, %zmm15
-; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm29, %xmm17
-; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm16 = xmm29[0],zero,xmm29[1],zero,xmm29[2],zero,xmm29[3],zero,xmm29[4],zero,xmm29[5],zero,xmm29[6],zero,xmm29[7],zero
-; AVX512BW-FCP-NEXT: vpermt2w %zmm17, %zmm27, %zmm16
-; AVX512BW-FCP-NEXT: vpshufb %xmm31, %xmm28, %xmm31
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm28[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm31, %zmm27, %zmm17
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm27
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm31 = ymm27[0],ymm7[0],ymm27[1],ymm7[1],ymm27[2],ymm7[2],ymm27[3],ymm7[3],ymm27[4],ymm7[4],ymm27[5],ymm7[5],ymm27[6],ymm7[6],ymm27[7],ymm7[7],ymm27[16],ymm7[16],ymm27[17],ymm7[17],ymm27[18],ymm7[18],ymm27[19],ymm7[19],ymm27[20],ymm7[20],ymm27[21],ymm7[21],ymm27[22],ymm7[22],ymm27[23],ymm7[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm30[8],xmm18[8],xmm30[9],xmm18[9],xmm30[10],xmm18[10],xmm30[11],xmm18[11],xmm30[12],xmm18[12],xmm30[13],xmm18[13],xmm30[14],xmm18[14],xmm30[15],xmm18[15]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm31, %zmm0, %zmm18
+; AVX512BW-FCP-NEXT: vpshufb %xmm29, %xmm22, %xmm16
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm22[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm16, %zmm28, %zmm15
+; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm26, %xmm19
+; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm16 = xmm26[0],zero,xmm26[1],zero,xmm26[2],zero,xmm26[3],zero,xmm26[4],zero,xmm26[5],zero,xmm26[6],zero,xmm26[7],zero
+; AVX512BW-FCP-NEXT: vpermt2w %zmm19, %zmm28, %zmm16
+; AVX512BW-FCP-NEXT: vpshufb %xmm29, %xmm27, %xmm29
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm27[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm29, %zmm28, %zmm19
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm28
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm5[0],ymm28[1],ymm5[1],ymm28[2],ymm5[2],ymm28[3],ymm5[3],ymm28[4],ymm5[4],ymm28[5],ymm5[5],ymm28[6],ymm5[6],ymm28[7],ymm5[7],ymm28[16],ymm5[16],ymm28[17],ymm5[17],ymm28[18],ymm5[18],ymm28[19],ymm5[19],ymm28[20],ymm5[20],ymm28[21],ymm5[21],ymm28[22],ymm5[22],ymm28[23],ymm5[23]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm25[8],xmm23[8],xmm25[9],xmm23[9],xmm25[10],xmm23[10],xmm25[11],xmm23[11],xmm25[12],xmm23[12],xmm25[13],xmm23[13],xmm25[14],xmm23[14],xmm25[15],xmm23[15]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm29, %zmm1, %zmm25
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm29
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm31[8],xmm30[8],xmm31[9],xmm30[9],xmm31[10],xmm30[10],xmm31[11],xmm30[11],xmm31[12],xmm30[12],xmm31[13],xmm30[13],xmm31[14],xmm30[14],xmm31[15],xmm30[15]
; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm30
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm24[8],xmm21[8],xmm24[9],xmm21[9],xmm24[10],xmm21[10],xmm24[11],xmm21[11],xmm24[12],xmm21[12],xmm24[13],xmm21[13],xmm24[14],xmm21[14],xmm24[15],xmm21[15]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm24 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm24, %zmm0, %zmm21
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %ymm31
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm30[0],ymm6[0],ymm30[1],ymm6[1],ymm30[2],ymm6[2],ymm30[3],ymm6[3],ymm30[4],ymm6[4],ymm30[5],ymm6[5],ymm30[6],ymm6[6],ymm30[7],ymm6[7],ymm30[16],ymm6[16],ymm30[17],ymm6[17],ymm30[18],ymm6[18],ymm30[19],ymm6[19],ymm30[20],ymm6[20],ymm30[21],ymm6[21],ymm30[22],ymm6[22],ymm30[23],ymm6[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm26 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm0, %zmm26, %zmm24
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm31 = ymm30[0],ymm29[0],ymm30[1],ymm29[1],ymm30[2],ymm29[2],ymm30[3],ymm29[3],ymm30[4],ymm29[4],ymm30[5],ymm29[5],ymm30[6],ymm29[6],ymm30[7],ymm29[7],ymm30[16],ymm29[16],ymm30[17],ymm29[17],ymm30[18],ymm29[18],ymm30[19],ymm29[19],ymm30[20],ymm29[20],ymm30[21],ymm29[21],ymm30[22],ymm29[22],ymm30[23],ymm29[23]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm31, %zmm2, %zmm23
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[16],ymm10[16],ymm11[17],ymm10[17],ymm11[18],ymm10[18],ymm11[19],ymm10[19],ymm11[20],ymm10[20],ymm11[21],ymm10[21],ymm11[22],ymm10[22],ymm11[23],ymm10[23]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm21[8],xmm17[8],xmm21[9],xmm17[9],xmm21[10],xmm17[10],xmm21[11],xmm17[11],xmm21[12],xmm17[12],xmm21[13],xmm17[13],xmm21[14],xmm17[14],xmm21[15],xmm17[15]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %ymm17
+; AVX512BW-FCP-NEXT: vpermt2w %zmm11, %zmm1, %zmm10
; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512BW-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm29[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm31, %ymm29
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm29, %zmm0, %zmm25
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[16],ymm9[16],ymm10[17],ymm9[17],ymm10[18],ymm9[18],ymm10[19],ymm9[19],ymm10[20],ymm9[20],ymm10[21],ymm9[21],ymm10[22],ymm9[22],ymm10[23],ymm9[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm22[8],xmm19[8],xmm22[9],xmm19[9],xmm22[10],xmm19[10],xmm22[11],xmm19[11],xmm22[12],xmm19[12],xmm22[13],xmm19[13],xmm22[14],xmm19[14],xmm22[15],xmm19[15]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %ymm22
-; AVX512BW-FCP-NEXT: vpermt2w %zmm10, %zmm26, %zmm9
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
-; AVX512BW-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm22, %ymm19
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm28[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm19, %zmm0, %zmm10
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm19
-; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm19, %ymm1
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm0, %zmm19
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm17, %ymm21
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm26[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm26 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm21, %zmm26, %zmm11
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm24[8],xmm18[8],xmm24[9],xmm18[9],xmm24[10],xmm18[10],xmm24[11],xmm18[11],xmm24[12],xmm18[12],xmm24[13],xmm18[13],xmm24[14],xmm18[14],xmm24[15],xmm18[15]
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %ymm21
+; AVX512BW-FCP-NEXT: vpermt2w %zmm7, %zmm2, %zmm6
+; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
+; AVX512BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm21, %ymm18
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm27[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm18, %zmm26, %zmm7
+; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm18
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm18, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm24
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm20[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm26, %zmm18
; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm1, %ymm1
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm23[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm0, %zmm20
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm0
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm1
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15],ymm5[24],ymm4[24],ymm5[25],ymm4[25],ymm5[26],ymm4[26],ymm5[27],ymm4[27],ymm5[28],ymm4[28],ymm5[29],ymm4[29],ymm5[30],ymm4[30],ymm5[31],ymm4[31]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm4
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm0
-; AVX512BW-FCP-NEXT: movl $613566756, %eax # imm = 0x24924924
-; AVX512BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm1[4,5,6,7,4,5,6,7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm2
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm22[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm26, %zmm20
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm24[4,5,6,7,4,5,6,7]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm22 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15]
+; AVX512BW-FCP-NEXT: vpshufb %zmm22, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512BW-FCP-NEXT: movl $-1840700270, %eax # imm = 0x92492492
; AVX512BW-FCP-NEXT: kmovd %eax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm4 {%k2}
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm0[4,5,6,7,4,5,6,7]
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[4,5,6,7,4,5,6,7]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512BW-FCP-NEXT: movabsq $-9076969306111049208, %rax # imm = 0x8208208208208208
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm4 {%k3}
-; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm23 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512BW-FCP-NEXT: vpshufb %ymm23, %ymm6, %ymm3
-; AVX512BW-FCP-NEXT: vpshufb %ymm23, %ymm30, %ymm23
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm23[0],ymm3[0],ymm23[1],ymm3[1],ymm23[2],ymm3[2],ymm23[3],ymm3[3],ymm23[4],ymm3[4],ymm23[5],ymm3[5],ymm23[6],ymm3[6],ymm23[7],ymm3[7],ymm23[16],ymm3[16],ymm23[17],ymm3[17],ymm23[18],ymm3[18],ymm23[19],ymm3[19],ymm23[20],ymm3[20],ymm23[21],ymm3[21],ymm23[22],ymm3[22],ymm23[23],ymm3[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm30[8],ymm6[8],ymm30[9],ymm6[9],ymm30[10],ymm6[10],ymm30[11],ymm6[11],ymm30[12],ymm6[12],ymm30[13],ymm6[13],ymm30[14],ymm6[14],ymm30[15],ymm6[15],ymm30[24],ymm6[24],ymm30[25],ymm6[25],ymm30[26],ymm6[26],ymm30[27],ymm6[27],ymm30[28],ymm6[28],ymm30[29],ymm6[29],ymm30[30],ymm6[30],ymm30[31],ymm6[31]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512BW-FCP-NEXT: vpermw %ymm6, %ymm23, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
+; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm26 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm29, %ymm1
+; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm30, %ymm26
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm26[0],ymm1[0],ymm26[1],ymm1[1],ymm26[2],ymm1[2],ymm26[3],ymm1[3],ymm26[4],ymm1[4],ymm26[5],ymm1[5],ymm26[6],ymm1[6],ymm26[7],ymm1[7],ymm26[16],ymm1[16],ymm26[17],ymm1[17],ymm26[18],ymm1[18],ymm26[19],ymm1[19],ymm26[20],ymm1[20],ymm26[21],ymm1[21],ymm26[22],ymm1[22],ymm26[23],ymm1[23]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm26 = ymm30[8],ymm29[8],ymm30[9],ymm29[9],ymm30[10],ymm29[10],ymm30[11],ymm29[11],ymm30[12],ymm29[12],ymm30[13],ymm29[13],ymm30[14],ymm29[14],ymm30[15],ymm29[15],ymm30[24],ymm29[24],ymm30[25],ymm29[25],ymm30[26],ymm29[26],ymm30[27],ymm29[27],ymm30[28],ymm29[28],ymm30[29],ymm29[29],ymm30[30],ymm29[30],ymm30[31],ymm29[31]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm27 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512BW-FCP-NEXT: vpermw %ymm26, %ymm27, %ymm26
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm26
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm28, %ymm3
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm26[0],ymm3[1],ymm26[1],ymm3[2],ymm26[2],ymm3[3],ymm26[3],ymm3[4],ymm26[4],ymm3[5],ymm26[5],ymm3[6],ymm26[6],ymm3[7],ymm26[7],ymm3[16],ymm26[16],ymm3[17],ymm26[17],ymm3[18],ymm26[18],ymm3[19],ymm26[19],ymm3[20],ymm26[20],ymm3[21],ymm26[21],ymm3[22],ymm26[22],ymm3[23],ymm26[23]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm28[8],ymm5[8],ymm28[9],ymm5[9],ymm28[10],ymm5[10],ymm28[11],ymm5[11],ymm28[12],ymm5[12],ymm28[13],ymm5[13],ymm28[14],ymm5[14],ymm28[15],ymm5[15],ymm28[24],ymm5[24],ymm28[25],ymm5[25],ymm28[26],ymm5[26],ymm28[27],ymm5[27],ymm28[28],ymm5[28],ymm28[29],ymm5[29],ymm28[30],ymm5[30],ymm28[31],ymm5[31]
+; AVX512BW-FCP-NEXT: vpermw %ymm5, %ymm4, %ymm4
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm7, %ymm6
-; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm27, %ymm2
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm6[0],ymm2[1],ymm6[1],ymm2[2],ymm6[2],ymm2[3],ymm6[3],ymm2[4],ymm6[4],ymm2[5],ymm6[5],ymm2[6],ymm6[6],ymm2[7],ymm6[7],ymm2[16],ymm6[16],ymm2[17],ymm6[17],ymm2[18],ymm6[18],ymm2[19],ymm6[19],ymm2[20],ymm6[20],ymm2[21],ymm6[21],ymm2[22],ymm6[22],ymm2[23],ymm6[23]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm27[8],ymm7[8],ymm27[9],ymm7[9],ymm27[10],ymm7[10],ymm27[11],ymm7[11],ymm27[12],ymm7[12],ymm27[13],ymm7[13],ymm27[14],ymm7[14],ymm27[15],ymm7[15],ymm27[24],ymm7[24],ymm27[25],ymm7[25],ymm27[26],ymm7[26],ymm27[27],ymm7[27],ymm27[28],ymm7[28],ymm27[29],ymm7[29],ymm27[30],ymm7[30],ymm27[31],ymm7[31]
-; AVX512BW-FCP-NEXT: vpermw %ymm6, %ymm5, %ymm5
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm1, %zmm1
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm24, %zmm1
+; AVX512BW-FCP-NEXT: vpshufb %zmm22, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm3 {%k2}
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm2 {%k2}
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm0
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,2,2,3,6,6,6,7]
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm2 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm11, %zmm8 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm9, %zmm8 {%k2}
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm12, %zmm8 {%k1}
; AVX512BW-FCP-NEXT: movabsq $585610922974906400, %rax # imm = 0x820820820820820
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm8 {%k3}
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm14, %zmm13 {%k2}
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm16, %zmm13 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm13 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm18, %zmm24 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm13 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm25, %zmm23 {%k1}
; AVX512BW-FCP-NEXT: movl $1227133513, %eax # imm = 0x49249249
; AVX512BW-FCP-NEXT: kmovd %eax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm25, %zmm24 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm11, %zmm23 {%k2}
; AVX512BW-FCP-NEXT: movabsq $2342443691899625602, %rax # imm = 0x2082082082082082
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm24 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm21, %zmm9 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqu16 %zmm19, %zmm9 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm9 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm23 {%k3}
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm10, %zmm6 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu16 %zmm18, %zmm6 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm6 {%k3}
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 256(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, 64(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, 64(%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, (%rax)
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 192(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 320(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%rax)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
@@ -6757,7 +6759,8 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: kmovd %r10d, %k1
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm9, %zmm0 {%k1}
; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm7[4,5,6,7,4,5,6,7]
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm9[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm14 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15]
+; AVX512DQ-BW-NEXT: vpshufb %zmm14, %zmm9, %zmm9
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492
; AVX512DQ-BW-NEXT: kmovd %r10d, %k2
@@ -6789,7 +6792,7 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm8, %zmm5 {%k1}
; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm13
; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm7 = zmm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512DQ-BW-NEXT: vpshufb %zmm14, %zmm7, %zmm7
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm7, %zmm5 {%k2}
; AVX512DQ-BW-NEXT: vmovdqa (%r9), %ymm12
@@ -6914,184 +6917,185 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf64:
; AVX512DQ-BW-FCP: # %bb.0:
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm9
-; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[16],ymm1[16],ymm2[17],ymm1[17],ymm2[18],ymm1[18],ymm2[19],ymm1[19],ymm2[20],ymm1[20],ymm2[21],ymm1[21],ymm2[22],ymm1[22],ymm2[23],ymm1[23]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm1[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15],ymm10[24],ymm9[24],ymm10[25],ymm9[25],ymm10[26],ymm9[26],ymm10[27],ymm9[27],ymm10[28],ymm9[28],ymm10[29],ymm9[29],ymm10[30],ymm9[30],ymm10[31],ymm9[31]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm3, %ymm0, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm4
-; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm18
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm21
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm21, %xmm8
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm30
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm24
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm24, %xmm11
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm24[0],xmm21[0],xmm24[1],xmm21[1],xmm24[2],xmm21[2],xmm24[3],xmm21[3],xmm24[4],xmm21[4],xmm24[5],xmm21[5],xmm24[6],xmm21[6],xmm24[7],xmm21[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm1 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm7
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[16],ymm0[16],ymm2[17],ymm0[17],ymm2[18],ymm0[18],ymm2[19],ymm0[19],ymm2[20],ymm0[20],ymm2[21],ymm0[21],ymm2[22],ymm0[22],ymm2[23],ymm0[23]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm1 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm3, %ymm1, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
+; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm4
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[16],ymm0[16],ymm4[17],ymm0[17],ymm4[18],ymm0[18],ymm4[19],ymm0[19],ymm4[20],ymm0[20],ymm4[21],ymm0[21],ymm4[22],ymm0[22],ymm4[23],ymm0[23]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11],ymm11[12],ymm10[12],ymm11[13],ymm10[13],ymm11[14],ymm10[14],ymm11[15],ymm10[15],ymm11[24],ymm10[24],ymm11[25],ymm10[25],ymm11[26],ymm10[26],ymm11[27],ymm10[27],ymm11[28],ymm10[28],ymm11[29],ymm10[29],ymm11[30],ymm10[30],ymm11[31],ymm10[31]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm8, %ymm4, %ymm8
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT: movl $613566756, %eax # imm = 0x24924924
+; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm0 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm23
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm17
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm3, %xmm17, %xmm8
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm25
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm21
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm3, %xmm21, %xmm9
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm21[0],xmm17[0],xmm21[1],xmm17[1],xmm21[2],xmm17[2],xmm21[3],xmm17[3],xmm21[4],xmm17[4],xmm21[5],xmm17[5],xmm21[6],xmm17[6],xmm21[7],xmm17[7]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm13 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm8, %zmm13, %zmm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm25
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm8, %zmm13, %zmm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm30
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm18
; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm15 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm19, %xmm8
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm26
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm22
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm22, %xmm12
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm18, %xmm8
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm31
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm24
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm24, %xmm12
; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm22[0],xmm19[0],xmm22[1],xmm19[1],xmm22[2],xmm19[2],xmm22[3],xmm19[3],xmm22[4],xmm19[4],xmm22[5],xmm19[5],xmm22[6],xmm19[6],xmm22[7],xmm19[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm24[0],xmm18[0],xmm24[1],xmm18[1],xmm24[2],xmm18[2],xmm24[3],xmm18[3],xmm24[4],xmm18[4],xmm24[5],xmm18[5],xmm24[6],xmm18[6],xmm24[7],xmm18[7]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm12, %zmm16, %zmm8
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %xmm29
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %xmm26
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm20
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm17 = [8,9,0,0,0,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm20, %xmm14
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm19 = [8,9,0,0,0,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm20, %xmm14
; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm12 = xmm20[0],zero,xmm20[1],zero,xmm20[2],zero,xmm20[3],zero,xmm20[4],zero,xmm20[5],zero,xmm20[6],zero,xmm20[7],zero
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm27 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm14, %zmm27, %zmm12
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %xmm28
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm23
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm18, %xmm14
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm30, %xmm31
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm31 = xmm31[0],xmm14[0],xmm31[1],xmm14[1],xmm31[2],xmm14[2],xmm31[3],xmm14[3],xmm31[4],xmm14[4],xmm31[5],xmm14[5],xmm31[6],xmm14[6],xmm31[7],xmm14[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm30[0],xmm18[0],xmm30[1],xmm18[1],xmm30[2],xmm18[2],xmm30[3],xmm18[3],xmm30[4],xmm18[4],xmm30[5],xmm18[5],xmm30[6],xmm18[6],xmm30[7],xmm18[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm31, %zmm13, %zmm14
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm31 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm25, %xmm13
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm26, %xmm15
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm14, %zmm28, %zmm12
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %xmm27
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm3, %xmm23, %xmm14
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm3, %xmm25, %xmm29
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm29 = xmm29[0],xmm14[0],xmm29[1],xmm14[1],xmm29[2],xmm14[2],xmm29[3],xmm14[3],xmm29[4],xmm14[4],xmm29[5],xmm14[5],xmm29[6],xmm14[6],xmm29[7],xmm14[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm25[0],xmm23[0],xmm25[1],xmm23[1],xmm25[2],xmm23[2],xmm25[3],xmm23[3],xmm25[4],xmm23[4],xmm25[5],xmm23[5],xmm25[6],xmm23[6],xmm25[7],xmm23[7]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm29, %zmm13, %zmm14
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm29 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm30, %xmm13
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm15, %xmm31, %xmm15
; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm15[8],xmm13[8],xmm15[9],xmm13[9],xmm15[10],xmm13[10],xmm15[11],xmm13[11],xmm15[12],xmm13[12],xmm15[13],xmm13[13],xmm15[14],xmm13[14],xmm15[15],xmm13[15]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm26[0],xmm25[0],xmm26[1],xmm25[1],xmm26[2],xmm25[2],xmm26[3],xmm25[3],xmm26[4],xmm25[4],xmm26[5],xmm25[5],xmm26[6],xmm25[6],xmm26[7],xmm25[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm31[0],xmm30[0],xmm31[1],xmm30[1],xmm31[2],xmm30[2],xmm31[3],xmm30[3],xmm31[4],xmm30[4],xmm31[5],xmm30[5],xmm31[6],xmm30[6],xmm31[7],xmm30[7]
; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm15, %zmm16, %zmm13
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm31, %xmm23, %xmm16
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm23[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm16, %zmm27, %zmm15
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm29, %xmm17
-; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm16 = xmm29[0],zero,xmm29[1],zero,xmm29[2],zero,xmm29[3],zero,xmm29[4],zero,xmm29[5],zero,xmm29[6],zero,xmm29[7],zero
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm17, %zmm27, %zmm16
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm31, %xmm28, %xmm31
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm28[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm31, %zmm27, %zmm17
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm27
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm31 = ymm27[0],ymm7[0],ymm27[1],ymm7[1],ymm27[2],ymm7[2],ymm27[3],ymm7[3],ymm27[4],ymm7[4],ymm27[5],ymm7[5],ymm27[6],ymm7[6],ymm27[7],ymm7[7],ymm27[16],ymm7[16],ymm27[17],ymm7[17],ymm27[18],ymm7[18],ymm27[19],ymm7[19],ymm27[20],ymm7[20],ymm27[21],ymm7[21],ymm27[22],ymm7[22],ymm27[23],ymm7[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm30[8],xmm18[8],xmm30[9],xmm18[9],xmm30[10],xmm18[10],xmm30[11],xmm18[11],xmm30[12],xmm18[12],xmm30[13],xmm18[13],xmm30[14],xmm18[14],xmm30[15],xmm18[15]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm31, %zmm0, %zmm18
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm29, %xmm22, %xmm16
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm22[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm16, %zmm28, %zmm15
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm26, %xmm19
+; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm16 = xmm26[0],zero,xmm26[1],zero,xmm26[2],zero,xmm26[3],zero,xmm26[4],zero,xmm26[5],zero,xmm26[6],zero,xmm26[7],zero
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm19, %zmm28, %zmm16
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm29, %xmm27, %xmm29
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm27[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm29, %zmm28, %zmm19
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm28
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm29 = ymm28[0],ymm5[0],ymm28[1],ymm5[1],ymm28[2],ymm5[2],ymm28[3],ymm5[3],ymm28[4],ymm5[4],ymm28[5],ymm5[5],ymm28[6],ymm5[6],ymm28[7],ymm5[7],ymm28[16],ymm5[16],ymm28[17],ymm5[17],ymm28[18],ymm5[18],ymm28[19],ymm5[19],ymm28[20],ymm5[20],ymm28[21],ymm5[21],ymm28[22],ymm5[22],ymm28[23],ymm5[23]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm25[8],xmm23[8],xmm25[9],xmm23[9],xmm25[10],xmm23[10],xmm25[11],xmm23[11],xmm25[12],xmm23[12],xmm25[13],xmm23[13],xmm25[14],xmm23[14],xmm25[15],xmm23[15]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm29, %zmm1, %zmm25
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm29
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm31[8],xmm30[8],xmm31[9],xmm30[9],xmm31[10],xmm30[10],xmm31[11],xmm30[11],xmm31[12],xmm30[12],xmm31[13],xmm30[13],xmm31[14],xmm30[14],xmm31[15],xmm30[15]
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm30
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm24[8],xmm21[8],xmm24[9],xmm21[9],xmm24[10],xmm21[10],xmm24[11],xmm21[11],xmm24[12],xmm21[12],xmm24[13],xmm21[13],xmm24[14],xmm21[14],xmm24[15],xmm21[15]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm24 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm24, %zmm0, %zmm21
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %ymm31
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm30[0],ymm6[0],ymm30[1],ymm6[1],ymm30[2],ymm6[2],ymm30[3],ymm6[3],ymm30[4],ymm6[4],ymm30[5],ymm6[5],ymm30[6],ymm6[6],ymm30[7],ymm6[7],ymm30[16],ymm6[16],ymm30[17],ymm6[17],ymm30[18],ymm6[18],ymm30[19],ymm6[19],ymm30[20],ymm6[20],ymm30[21],ymm6[21],ymm30[22],ymm6[22],ymm30[23],ymm6[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm26 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm0, %zmm26, %zmm24
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm31 = ymm30[0],ymm29[0],ymm30[1],ymm29[1],ymm30[2],ymm29[2],ymm30[3],ymm29[3],ymm30[4],ymm29[4],ymm30[5],ymm29[5],ymm30[6],ymm29[6],ymm30[7],ymm29[7],ymm30[16],ymm29[16],ymm30[17],ymm29[17],ymm30[18],ymm29[18],ymm30[19],ymm29[19],ymm30[20],ymm29[20],ymm30[21],ymm29[21],ymm30[22],ymm29[22],ymm30[23],ymm29[23]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm31, %zmm2, %zmm23
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[16],ymm10[16],ymm11[17],ymm10[17],ymm11[18],ymm10[18],ymm11[19],ymm10[19],ymm11[20],ymm10[20],ymm11[21],ymm10[21],ymm11[22],ymm10[22],ymm11[23],ymm10[23]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm21[8],xmm17[8],xmm21[9],xmm17[9],xmm21[10],xmm17[10],xmm21[11],xmm17[11],xmm21[12],xmm17[12],xmm21[13],xmm17[13],xmm21[14],xmm17[14],xmm21[15],xmm17[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %ymm17
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm11, %zmm1, %zmm10
; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512DQ-BW-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm29[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm31, %ymm29
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm0 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm29, %zmm0, %zmm25
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[16],ymm9[16],ymm10[17],ymm9[17],ymm10[18],ymm9[18],ymm10[19],ymm9[19],ymm10[20],ymm9[20],ymm10[21],ymm9[21],ymm10[22],ymm9[22],ymm10[23],ymm9[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm22[8],xmm19[8],xmm22[9],xmm19[9],xmm22[10],xmm19[10],xmm22[11],xmm19[11],xmm22[12],xmm19[12],xmm22[13],xmm19[13],xmm22[14],xmm19[14],xmm22[15],xmm19[15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %ymm22
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm10, %zmm26, %zmm9
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
-; AVX512DQ-BW-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm22, %ymm19
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm28[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm19, %zmm0, %zmm10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm19
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm19, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm0, %zmm19
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm17, %ymm21
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm26[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm26 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm21, %zmm26, %zmm11
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm24[8],xmm18[8],xmm24[9],xmm18[9],xmm24[10],xmm18[10],xmm24[11],xmm18[11],xmm24[12],xmm18[12],xmm24[13],xmm18[13],xmm24[14],xmm18[14],xmm24[15],xmm18[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %ymm21
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm7, %zmm2, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
+; AVX512DQ-BW-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm21, %ymm18
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm27[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm18, %zmm26, %zmm7
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm18
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm18, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm24
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm20[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm26, %zmm18
; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm1, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm23[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm0, %zmm20
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm0
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11],ymm5[12],ymm4[12],ymm5[13],ymm4[13],ymm5[14],ymm4[14],ymm5[15],ymm4[15],ymm5[24],ymm4[24],ymm5[25],ymm4[25],ymm5[26],ymm4[26],ymm5[27],ymm4[27],ymm5[28],ymm4[28],ymm5[29],ymm4[29],ymm5[30],ymm4[30],ymm5[31],ymm4[31]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm4, %ymm5, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm0
-; AVX512DQ-BW-FCP-NEXT: movl $613566756, %eax # imm = 0x24924924
-; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm1[4,5,6,7,4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm22[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm26, %zmm20
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm24[4,5,6,7,4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm22 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm22, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-FCP-NEXT: movl $-1840700270, %eax # imm = 0x92492492
; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm4 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm0[4,5,6,7,4,5,6,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[4,5,6,7,4,5,6,7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
; AVX512DQ-BW-FCP-NEXT: movabsq $-9076969306111049208, %rax # imm = 0x8208208208208208
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm4 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm23 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm23, %ymm6, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm23, %ymm30, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm23[0],ymm3[0],ymm23[1],ymm3[1],ymm23[2],ymm3[2],ymm23[3],ymm3[3],ymm23[4],ymm3[4],ymm23[5],ymm3[5],ymm23[6],ymm3[6],ymm23[7],ymm3[7],ymm23[16],ymm3[16],ymm23[17],ymm3[17],ymm23[18],ymm3[18],ymm23[19],ymm3[19],ymm23[20],ymm3[20],ymm23[21],ymm3[21],ymm23[22],ymm3[22],ymm23[23],ymm3[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm30[8],ymm6[8],ymm30[9],ymm6[9],ymm30[10],ymm6[10],ymm30[11],ymm6[11],ymm30[12],ymm6[12],ymm30[13],ymm6[13],ymm30[14],ymm6[14],ymm30[15],ymm6[15],ymm30[24],ymm6[24],ymm30[25],ymm6[25],ymm30[26],ymm6[26],ymm30[27],ymm6[27],ymm30[28],ymm6[28],ymm30[29],ymm6[29],ymm30[30],ymm6[30],ymm30[31],ymm6[31]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm6, %ymm23, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm26 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm29, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm30, %ymm26
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm26[0],ymm1[0],ymm26[1],ymm1[1],ymm26[2],ymm1[2],ymm26[3],ymm1[3],ymm26[4],ymm1[4],ymm26[5],ymm1[5],ymm26[6],ymm1[6],ymm26[7],ymm1[7],ymm26[16],ymm1[16],ymm26[17],ymm1[17],ymm26[18],ymm1[18],ymm26[19],ymm1[19],ymm26[20],ymm1[20],ymm26[21],ymm1[21],ymm26[22],ymm1[22],ymm26[23],ymm1[23]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm26 = ymm30[8],ymm29[8],ymm30[9],ymm29[9],ymm30[10],ymm29[10],ymm30[11],ymm29[11],ymm30[12],ymm29[12],ymm30[13],ymm29[13],ymm30[14],ymm29[14],ymm30[15],ymm29[15],ymm30[24],ymm29[24],ymm30[25],ymm29[25],ymm30[26],ymm29[26],ymm30[27],ymm29[27],ymm30[28],ymm29[28],ymm30[29],ymm29[29],ymm30[30],ymm29[30],ymm30[31],ymm29[31]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm27 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm26, %ymm27, %ymm26
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm26
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm28, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm26[0],ymm3[1],ymm26[1],ymm3[2],ymm26[2],ymm3[3],ymm26[3],ymm3[4],ymm26[4],ymm3[5],ymm26[5],ymm3[6],ymm26[6],ymm3[7],ymm26[7],ymm3[16],ymm26[16],ymm3[17],ymm26[17],ymm3[18],ymm26[18],ymm3[19],ymm26[19],ymm3[20],ymm26[20],ymm3[21],ymm26[21],ymm3[22],ymm26[22],ymm3[23],ymm26[23]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm28[8],ymm5[8],ymm28[9],ymm5[9],ymm28[10],ymm5[10],ymm28[11],ymm5[11],ymm28[12],ymm5[12],ymm28[13],ymm5[13],ymm28[14],ymm5[14],ymm28[15],ymm5[15],ymm28[24],ymm5[24],ymm28[25],ymm5[25],ymm28[26],ymm5[26],ymm28[27],ymm5[27],ymm28[28],ymm5[28],ymm28[29],ymm5[29],ymm28[30],ymm5[30],ymm28[31],ymm5[31]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm5, %ymm4, %ymm4
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm7, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm27, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm6[0],ymm2[1],ymm6[1],ymm2[2],ymm6[2],ymm2[3],ymm6[3],ymm2[4],ymm6[4],ymm2[5],ymm6[5],ymm2[6],ymm6[6],ymm2[7],ymm6[7],ymm2[16],ymm6[16],ymm2[17],ymm6[17],ymm2[18],ymm6[18],ymm2[19],ymm6[19],ymm2[20],ymm6[20],ymm2[21],ymm6[21],ymm2[22],ymm6[22],ymm2[23],ymm6[23]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm27[8],ymm7[8],ymm27[9],ymm7[9],ymm27[10],ymm7[10],ymm27[11],ymm7[11],ymm27[12],ymm7[12],ymm27[13],ymm7[13],ymm27[14],ymm7[14],ymm27[15],ymm7[15],ymm27[24],ymm7[24],ymm27[25],ymm7[25],ymm27[26],ymm7[26],ymm27[27],ymm7[27],ymm27[28],ymm7[28],ymm27[29],ymm7[29],ymm27[30],ymm7[30],ymm27[31],ymm7[31]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm6, %ymm5, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm3, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm1, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63,u]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm24, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm22, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm3 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm2 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm0
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63]
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,2,2,3,6,6,6,7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm2 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm11, %zmm8 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm3 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm9, %zmm8 {%k2}
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm12, %zmm8 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $585610922974906400, %rax # imm = 0x820820820820820
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm8 {%k3}
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm14, %zmm13 {%k2}
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm16, %zmm13 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm13 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm18, %zmm24 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm13 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm25, %zmm23 {%k1}
; AVX512DQ-BW-FCP-NEXT: movl $1227133513, %eax # imm = 0x49249249
; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm25, %zmm24 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm11, %zmm23 {%k2}
; AVX512DQ-BW-FCP-NEXT: movabsq $2342443691899625602, %rax # imm = 0x2082082082082082
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm24 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm21, %zmm9 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm19, %zmm9 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm9 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm7, %zmm23 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm10, %zmm6 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm18, %zmm6 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm20, %zmm6 {%k3}
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 256(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, 64(%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, (%rax)
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 192(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%rax)
; 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 aadb8b7635636..363131b1a1045 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
@@ -4098,132 +4098,134 @@ 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), %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: 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: vporq %zmm7, %zmm8, %zmm7
-; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; 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: 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: vporq %zmm8, %zmm9, %zmm8
-; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
-; 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: 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: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (mem & (zmm7 ^ zmm10))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm8))
-; 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: vshufi64x2 {{.*#+}} zmm8 = zmm11[0,1,0,1],zmm8[0,1,0,1]
-; 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: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm11[0,1,0,1]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm8))
-; 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: vshufi64x2 {{.*#+}} zmm17 = zmm13[0,1,0,1],zmm8[0,1,0,1]
-; 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: vmovdqa (%rdi), %ymm6
+; AVX512-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512-NEXT: vmovdqa64 (%rcx), %ymm19
+; AVX512-NEXT: vmovdqa (%r8), %ymm4
+; AVX512-NEXT: vmovdqa (%r9), %ymm5
+; AVX512-NEXT: vmovdqa64 (%r10), %ymm18
+; AVX512-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[u],zero,xmm0[7],zero,xmm0[5,u,u,u],zero,xmm0[8],zero,xmm0[6,u,u,u],zero
+; AVX512-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = xmm9[u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u,u,9]
+; AVX512-NEXT: vpor %xmm7, %xmm10, %xmm7
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3],xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm20
+; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm10[0,1,0,1],zmm7[0,1,0,1]
+; AVX512-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512-NEXT: vpshufb {{.*#+}} xmm10 = xmm13[u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6,u,u]
+; AVX512-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero,xmm14[u,u]
+; AVX512-NEXT: vpor %xmm10, %xmm11, %xmm10
+; AVX512-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]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm11[0,1,0,1],zmm10[0,1,0,1]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm7))
+; AVX512-NEXT: vmovdqa (%r9), %xmm10
+; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm10[4,u,u,u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6]
+; AVX512-NEXT: vmovdqa (%r8), %xmm11
+; AVX512-NEXT: vpshufb {{.*#+}} xmm12 = xmm11[4],zero,xmm11[u,u,u,7],zero,xmm11[5],zero,xmm11[u,u,u,8],zero,xmm11[6],zero
+; AVX512-NEXT: vpor %xmm7, %xmm12, %xmm7
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm7[0,1,0,1]
+; AVX512-NEXT: vmovdqa (%r10), %xmm12
+; AVX512-NEXT: vpshufb {{.*#+}} xmm7 = xmm12[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm12[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: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm17))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm16))
-; 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: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
+; AVX512-NEXT: vpermq {{.*#+}} zmm17 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm16))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm15))
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27],zero,ymm6[25]
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm7
+; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[18],zero,zero,zero,zero,ymm7[21],zero,ymm7[19],zero,zero,zero,zero,ymm7[22],zero,ymm7[20]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm15[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = 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 = 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: vshufi64x2 {{.*#+}} zmm1 = zmm1[2,3,2,3],zmm15[2,3,2,3]
+; AVX512-NEXT: vporq %zmm0, %zmm1, %zmm0
+; AVX512-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
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm16
+; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = 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
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm15[2,3,2,3],zmm1[2,3,2,3]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = 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 {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm6[19],zero,ymm6[21,20,21,22],zero,ymm6[20],zero,ymm6[22,23]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[2,3,2,3],zmm15[2,3,2,3]
+; AVX512-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; 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[25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[20],zero,ymm5[18],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22]
+; AVX512-NEXT: vmovdqa64 %ymm5, %ymm19
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-NEXT: vmovdqa %ymm4, %ymm5
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm4[23],zero,ymm4[23,24,25,26],zero,ymm4[24],zero,ymm4[30,31]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm4[18],zero,ymm4[20,21,20,21],zero,ymm4[19],zero,ymm4[19,20,21,22],zero
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm15[2,3,2,3],zmm2[2,3,2,3]
+; AVX512-NEXT: vporq %zmm0, %zmm2, %zmm0
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm4
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = 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 {{.*#+}} ymm7 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512-NEXT: vpermi2d %zmm2, %zmm7, %zmm15
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm0))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm1))
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14,u,u],zero,zero,zero,zero,ymm3[15,u,u],zero,zero,zero,zero,ymm3[16,u,u],zero,zero,zero,zero,ymm3[17,u,u],zero,zero,zero,zero,ymm3[18]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[0,1,14],zero,ymm6[u,u,0,1,14,15],zero,ymm6[u,u,13,2,3,16],zero,ymm6[u,u,28,29,16,17],zero,ymm6[u,u,19,28,29,18],zero
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
-; 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: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm14[8],xmm13[9],xmm14[9],xmm13[10],xmm14[10],xmm13[11],xmm14[11],xmm13[12],xmm14[12],xmm13[13],xmm14[13],xmm13[14],xmm14[14],xmm13[15],xmm14[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 = 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: vmovdqa64 %ymm16, %ymm13
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u],zero,ymm13[14,u,u,u,u,u],zero,ymm13[15,u,u,u,u,u],zero,ymm13[16,u,u,u,u,u],zero,ymm13[17,u,u,u,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,14],zero,ymm8[u,u,u,u,u,15],zero,ymm8[u,u,u,u,u,16],zero,ymm8[u,u,u,u,u,17],zero,ymm8[u,u,u,u,u]
+; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa64 %xmm20, %xmm2
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm9[8],xmm2[9],xmm9[9],xmm2[10],xmm9[10],xmm2[11],xmm9[11],xmm2[12],xmm9[12],xmm2[13],xmm9[13],xmm2[14],xmm9[14],xmm2[15],xmm9[15]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm14
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u],zero,zero,zero,zero,ymm14[14],zero,ymm14[u],zero,zero,zero,zero,ymm14[15],zero,ymm14[u],zero,zero,zero,zero,ymm14[16],zero,ymm14[u],zero,zero,zero,zero,ymm14[17],zero,ymm14[u],zero,zero
-; 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 {{.*#+}} zmm9 = zmm9 | (zmm0 & mem)
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm1))
-; 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: vpshufb {{.*#+}} ymm0 = zero,ymm5[u],zero,zero,zero,zero,ymm5[14],zero,ymm5[u],zero,zero,zero,zero,ymm5[15],zero,ymm5[u],zero,zero,zero,zero,ymm5[16],zero,ymm5[u],zero,zero,zero,zero,ymm5[17],zero,ymm5[u],zero,zero
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm9
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm9[13,u,u,u,u,u],zero,ymm9[14,u,u,u,u,u],zero,ymm9[15,u,u,u,u,u],zero,ymm9[16,u,u,u,u,u],zero,ymm9[17,u,u,u]
+; AVX512-NEXT: vpor %ymm0, %ymm2, %ymm0
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm12[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = 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, %ymm7, %zmm2, %zmm2
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm0 & mem)
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm13[30],zero,ymm13[28,u,u,u],zero,ymm13[31],zero,ymm13[29,u]
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; 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: vpshufb {{.*#+}} ymm1 = 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
+; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = 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]
; AVX512-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm0))
-; 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: vpshufb {{.*#+}} ymm0 = 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 {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm9[27,u,u,u],zero,ymm9[30],zero,ymm9[28,u,u,u],zero,ymm9[31],zero
+; AVX512-NEXT: vpor %ymm0, %ymm3, %ymm0
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; 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 {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm0))
-; AVX512-NEXT: vpternlogq {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm1))
-; 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: vpshufb {{.*#+}} ymm3 = 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: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm0))
+; AVX512-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm1))
+; AVX512-NEXT: vmovdqa %ymm3, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm15, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm17, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -4238,38 +4240,35 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm2
; AVX512-FCP-NEXT: vmovdqa (%r10), %ymm0
-; AVX512-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = 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: vpshufb {{.*#+}} ymm8 = 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: vshufi64x2 {{.*#+}} zmm7 = zmm8[2,3,2,3],zmm7[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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: 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: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
; AVX512-FCP-NEXT: vporq %zmm7, %zmm8, %zmm7
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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: 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: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: 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: vshufi64x2 {{.*#+}} zmm9 = zmm10[2,3,2,3],zmm9[2,3,2,3]
; AVX512-FCP-NEXT: vporq %zmm8, %zmm9, %zmm8
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm10 = [0,21,20,0,21,0,20,0,4,5,0,7,0,5,0,7]
-; AVX512-FCP-NEXT: vpermi2d %zmm7, %zmm9, %zmm10
-; AVX512-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512-FCP-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-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: vmovdqa64 %ymm1, %ymm19
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (mem & (zmm7 ^ zmm10))
+; AVX512-FCP-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
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: vshufi64x2 {{.*#+}} zmm7 = zmm9[2,3,2,3],zmm7[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: 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: vmovdqa64 %ymm1, %ymm18
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[2,3,2,3],zmm9[2,3,2,3]
+; AVX512-FCP-NEXT: vporq %zmm7, %zmm9, %zmm9
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm10 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = 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-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,21,20,0,21,0,20,0,4,5,0,7,0,5,0,7]
+; AVX512-FCP-NEXT: vpermi2d %zmm10, %zmm11, %zmm7
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm9))
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm8))
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm9
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm9[u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u,u],zero
@@ -4318,7 +4317,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm12
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm12
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm12[u],zero,zero,zero,zero,ymm12[14],zero,ymm12[u],zero,zero,zero,zero,ymm12[15],zero,ymm12[u],zero,zero,zero,zero,ymm12[16],zero,ymm12[u],zero,zero,zero,zero,ymm12[17],zero,ymm12[u],zero,zero
; AVX512-FCP-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-FCP-NEXT: vpor %ymm0, %ymm9, %ymm0
@@ -4331,7 +4330,7 @@ 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 %ymm9, %ymm10, %ymm9
; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm11
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm11
; AVX512-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm0 & mem)
@@ -4364,132 +4363,134 @@ 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), %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: 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: vporq %zmm7, %zmm8, %zmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; 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: 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: vporq %zmm8, %zmm9, %zmm8
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
-; 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: 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: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512DQ-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (mem & (zmm7 ^ zmm10))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm8))
-; 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: vshufi64x2 {{.*#+}} zmm8 = zmm11[0,1,0,1],zmm8[0,1,0,1]
-; 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: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm11[0,1,0,1]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm8))
-; 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: vshufi64x2 {{.*#+}} zmm17 = zmm13[0,1,0,1],zmm8[0,1,0,1]
-; 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: vmovdqa (%rdi), %ymm6
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512DQ-NEXT: vmovdqa64 (%rcx), %ymm19
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm4
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm5
+; AVX512DQ-NEXT: vmovdqa64 (%r10), %ymm18
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[u],zero,xmm0[7],zero,xmm0[5,u,u,u],zero,xmm0[8],zero,xmm0[6,u,u,u],zero
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = xmm9[u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u,u,9]
+; AVX512DQ-NEXT: vpor %xmm7, %xmm10, %xmm7
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3],xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm20
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm10[0,1,0,1],zmm7[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm10 = xmm13[u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6,u,u]
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero,xmm14[u,u]
+; AVX512DQ-NEXT: vpor %xmm10, %xmm11, %xmm10
+; AVX512DQ-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-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm11[0,1,0,1],zmm10[0,1,0,1]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm7))
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm10
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm10[4,u,u,u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6]
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm11
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm12 = xmm11[4],zero,xmm11[u,u,u,7],zero,xmm11[5],zero,xmm11[u,u,u,8],zero,xmm11[6],zero
+; AVX512DQ-NEXT: vpor %xmm7, %xmm12, %xmm7
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm7[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa (%r10), %xmm12
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm7 = xmm12[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm12[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: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm17))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm16))
-; 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: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm17 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm16))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm15))
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero,zero,zero,ymm6[27],zero,ymm6[25]
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm7
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[18],zero,zero,zero,zero,ymm7[21],zero,ymm7[19],zero,zero,zero,zero,ymm7[22],zero,ymm7[20]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm15[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = 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 = 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: vshufi64x2 {{.*#+}} zmm1 = zmm1[2,3,2,3],zmm15[2,3,2,3]
+; AVX512DQ-NEXT: vporq %zmm0, %zmm1, %zmm0
+; AVX512DQ-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
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm16
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = 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
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm15[2,3,2,3],zmm1[2,3,2,3]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = 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 {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm6[19],zero,ymm6[21,20,21,22],zero,ymm6[20],zero,ymm6[22,23]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[2,3,2,3],zmm15[2,3,2,3]
+; AVX512DQ-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
+; 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[25],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[20],zero,ymm5[18],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22]
+; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm19
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqa %ymm4, %ymm5
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm4[23],zero,ymm4[23,24,25,26],zero,ymm4[24],zero,ymm4[30,31]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm4[18],zero,ymm4[20,21,20,21],zero,ymm4[19],zero,ymm4[19,20,21,22],zero
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm15[2,3,2,3],zmm2[2,3,2,3]
+; AVX512DQ-NEXT: vporq %zmm0, %zmm2, %zmm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm4
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = 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 {{.*#+}} ymm7 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm7, %zmm15
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm0))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm1))
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14,u,u],zero,zero,zero,zero,ymm3[15,u,u],zero,zero,zero,zero,ymm3[16,u,u],zero,zero,zero,zero,ymm3[17,u,u],zero,zero,zero,zero,ymm3[18]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[0,1,14],zero,ymm6[u,u,0,1,14,15],zero,ymm6[u,u,13,2,3,16],zero,ymm6[u,u,28,29,16,17],zero,ymm6[u,u,19,28,29,18],zero
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
-; 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: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm14[8],xmm13[9],xmm14[9],xmm13[10],xmm14[10],xmm13[11],xmm14[11],xmm13[12],xmm14[12],xmm13[13],xmm14[13],xmm13[14],xmm14[14],xmm13[15],xmm14[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 = 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: vmovdqa64 %ymm16, %ymm13
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u],zero,ymm13[14,u,u,u,u,u],zero,ymm13[15,u,u,u,u,u],zero,ymm13[16,u,u,u,u,u],zero,ymm13[17,u,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,14],zero,ymm8[u,u,u,u,u,15],zero,ymm8[u,u,u,u,u,16],zero,ymm8[u,u,u,u,u,17],zero,ymm8[u,u,u,u,u]
+; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm2
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm9[8],xmm2[9],xmm9[9],xmm2[10],xmm9[10],xmm2[11],xmm9[11],xmm2[12],xmm9[12],xmm2[13],xmm9[13],xmm2[14],xmm9[14],xmm2[15],xmm9[15]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm14
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u],zero,zero,zero,zero,ymm14[14],zero,ymm14[u],zero,zero,zero,zero,ymm14[15],zero,ymm14[u],zero,zero,zero,zero,ymm14[16],zero,ymm14[u],zero,zero,zero,zero,ymm14[17],zero,ymm14[u],zero,zero
-; 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 {{.*#+}} zmm9 = zmm9 | (zmm0 & mem)
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm1))
-; 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: vpshufb {{.*#+}} ymm0 = zero,ymm5[u],zero,zero,zero,zero,ymm5[14],zero,ymm5[u],zero,zero,zero,zero,ymm5[15],zero,ymm5[u],zero,zero,zero,zero,ymm5[16],zero,ymm5[u],zero,zero,zero,zero,ymm5[17],zero,ymm5[u],zero,zero
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm9
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm9[13,u,u,u,u,u],zero,ymm9[14,u,u,u,u,u],zero,ymm9[15,u,u,u,u,u],zero,ymm9[16,u,u,u,u,u],zero,ymm9[17,u,u,u]
+; AVX512DQ-NEXT: vpor %ymm0, %ymm2, %ymm0
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm12[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = 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, %ymm7, %zmm2, %zmm2
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm0 & mem)
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm13[30],zero,ymm13[28,u,u,u],zero,ymm13[31],zero,ymm13[29,u]
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; 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: vpshufb {{.*#+}} ymm1 = 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-NEXT: vpshufb {{.*#+}} ymm3 = 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-NEXT: vpor %ymm1, %ymm3, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm0))
-; 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: vpshufb {{.*#+}} ymm0 = 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 {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm9[27,u,u,u],zero,ymm9[30],zero,ymm9[28,u,u,u],zero,ymm9[31],zero
+; AVX512DQ-NEXT: vpor %ymm0, %ymm3, %ymm0
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; 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 {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm0))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm1))
-; 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: vpshufb {{.*#+}} ymm3 = 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: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm0))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (mem & (ymm3 ^ ymm1))
+; AVX512DQ-NEXT: vmovdqa %ymm3, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm15, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm17, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 64(%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
@@ -4504,38 +4505,35 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa (%r10), %ymm0
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = 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: vpshufb {{.*#+}} ymm8 = 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: vshufi64x2 {{.*#+}} zmm7 = zmm8[2,3,2,3],zmm7[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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: 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: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm8, %zmm7
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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: 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: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: 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: vshufi64x2 {{.*#+}} zmm9 = zmm10[2,3,2,3],zmm9[2,3,2,3]
; AVX512DQ-FCP-NEXT: vporq %zmm8, %zmm9, %zmm8
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm10 = [0,21,20,0,21,0,20,0,4,5,0,7,0,5,0,7]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm7, %zmm9, %zmm10
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512DQ-FCP-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-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: vmovdqa64 %ymm1, %ymm19
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (mem & (zmm7 ^ zmm10))
+; AVX512DQ-FCP-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
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: vshufi64x2 {{.*#+}} zmm7 = zmm9[2,3,2,3],zmm7[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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: 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: vmovdqa64 %ymm1, %ymm18
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[2,3,2,3],zmm9[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm10 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = 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-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,21,20,0,21,0,20,0,4,5,0,7,0,5,0,7]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm10, %zmm11, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm9))
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm8))
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm9
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm9[u],zero,xmm9[7],zero,xmm9[5,u,u,u],zero,xmm9[8],zero,xmm9[6,u,u,u],zero
@@ -4584,7 +4582,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm12
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm12[u],zero,zero,zero,zero,ymm12[14],zero,ymm12[u],zero,zero,zero,zero,ymm12[15],zero,ymm12[u],zero,zero,zero,zero,ymm12[16],zero,ymm12[u],zero,zero,zero,zero,ymm12[17],zero,ymm12[u],zero,zero
; AVX512DQ-FCP-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-FCP-NEXT: vpor %ymm0, %ymm9, %ymm0
@@ -4597,7 +4595,7 @@ 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 %ymm9, %ymm10, %ymm9
; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm11
; AVX512DQ-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm0 & mem)
@@ -8399,2727 +8397,2585 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-LABEL: store_i8_stride7_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $1496, %rsp # imm = 0x5D8
-; AVX512-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [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 %ymm1, %ymm3, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512-NEXT: subq $1256, %rsp # imm = 0x4E8
+; AVX512-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm5 = [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 %ymm5, %ymm9, %ymm0
; AVX512-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [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 %ymm2, %ymm10, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [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 %ymm4, %ymm10, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %ymm8
+; AVX512-NEXT: vmovdqa (%rcx), %ymm11
; 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, %ymm8, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512-NEXT: vmovdqa (%rdx), %ymm7
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm7, %ymm1
-; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%r8), %ymm6
-; AVX512-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]
-; AVX512-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm24
-; AVX512-NEXT: vmovdqa (%r9), %ymm5
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm5, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512-NEXT: vmovdqa (%rdx), %ymm13
+; 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, %ymm13, %ymm2
+; AVX512-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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: # ymm14 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm14, %ymm3, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [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 %ymm0, %ymm10, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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: # ymm15 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm15, %ymm7, %ymm1
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm8, %ymm3
-; AVX512-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm6, %ymm1
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: vmovdqa (%r8), %ymm14
+; 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, %ymm14, %ymm0
+; AVX512-NEXT: vmovdqa (%r9), %ymm15
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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: # ymm6 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm6, %ymm15, %ymm7
+; AVX512-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm8, %ymm9, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm9, %ymm21
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm10, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm10, %ymm18
+; AVX512-NEXT: vporq %ymm0, %ymm7, %ymm31
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm9, %ymm5, %ymm13
-; AVX512-NEXT: vpor %ymm1, %ymm13, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm13
-; AVX512-NEXT: vpshufb %ymm14, %ymm13, %ymm1
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm14
-; AVX512-NEXT: vpshufb %ymm0, %ymm14, %ymm0
-; AVX512-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: vpshufb %ymm9, %ymm13, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm13, %ymm19
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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: # ymm10 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm10, %ymm11, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm11, %ymm16
+; AVX512-NEXT: vpor %ymm0, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero,zero,zero,ymm14[27],zero,ymm14[25]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm13[23,u,u,u],zero,ymm13[26],zero,ymm13[24,u,u,u],zero,ymm13[27],zero
-; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; 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, %ymm14, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm14, %ymm17
+; 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, %ymm15, %ymm13
+; AVX512-NEXT: vmovdqa64 %ymm15, %ymm24
+; AVX512-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm7
+; AVX512-NEXT: vpshufb %ymm8, %ymm7, %ymm13
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm8
+; AVX512-NEXT: vpshufb %ymm12, %ymm8, %ymm12
+; AVX512-NEXT: vpor %ymm13, %ymm12, %ymm12
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero,zero,zero,ymm8[27],zero,ymm8[25]
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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-NEXT: # ymm14 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm14, %ymm7, %ymm13
+; AVX512-NEXT: vmovdqa64 %ymm14, %ymm25
+; AVX512-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512-NEXT: vpshufb %ymm9, %ymm12, %ymm9
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512-NEXT: vpshufb %ymm10, %ymm13, %ymm10
+; AVX512-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm13[25],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm12[23],zero,ymm12[21,22,23,26],zero,ymm12[24],zero,ymm12[28,29,26,27]
+; AVX512-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm9
+; AVX512-NEXT: vpshufb %ymm0, %ymm9, %ymm0
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512-NEXT: vpshufb %ymm11, %ymm10, %ymm11
+; AVX512-NEXT: vpor %ymm0, %ymm11, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm10[25],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm11 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm9[23],zero,ymm9[23,24,25,26],zero,ymm9[24],zero,ymm9[30,31]
+; AVX512-NEXT: vpor %ymm0, %ymm11, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512-NEXT: vpshufb %ymm15, %ymm1, %ymm15
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512-NEXT: vpshufb %ymm2, %ymm0, %ymm2
-; AVX512-NEXT: vpor %ymm2, %ymm15, %ymm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[25],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm12, %ymm1, %ymm15
-; AVX512-NEXT: vmovdqa64 %ymm12, %ymm22
-; AVX512-NEXT: vpor %ymm2, %ymm15, %ymm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm15
-; AVX512-NEXT: vpshufb %ymm3, %ymm15, %ymm3
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX512-NEXT: vpor %ymm3, %ymm9, %ymm3
-; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm9, %ymm2, %ymm3
-; AVX512-NEXT: vmovdqa64 %ymm9, %ymm21
-; AVX512-NEXT: vpshufb {{.*#+}} ymm9 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm15[23],zero,ymm15[23,24,25,26],zero,ymm15[24],zero,ymm15[30,31]
-; AVX512-NEXT: vpor %ymm3, %ymm9, %ymm3
-; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm9[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 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm4, %ymm9, %ymm12
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm20
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm12[2,3,2,3],zmm3[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512-NEXT: vpshufb %ymm3, %ymm0, %ymm3
-; AVX512-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512-NEXT: vpor %ymm3, %ymm11, %ymm3
-; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm4, %ymm0, %ymm0
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm9, %ymm1, %ymm1
+; AVX512-NEXT: vmovdqa 32(%rax), %ymm14
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[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 {{.*#+}} ymm11 = 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]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm11[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm0
+; AVX512-NEXT: vpshufb %ymm3, %ymm12, %ymm1
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[18],zero,zero,zero,zero,ymm13[21],zero,ymm13[19],zero,zero,zero,zero,ymm13[22],zero,ymm13[20]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm12[18,19,20,21],zero,ymm12[19],zero,ymm12[25,26,27,22],zero,ymm12[20],zero
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm16, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512-NEXT: vpshufb %ymm5, %ymm7, %ymm0
+; AVX512-NEXT: vpshufb %ymm4, %ymm8, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm26
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm14, %ymm1
+; AVX512-NEXT: vpshufb %ymm1, %ymm7, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm8, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm15, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[20],zero,ymm10[18],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22]
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm8, %ymm9, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm15, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vpshufb %ymm2, %ymm9, %ymm0
+; AVX512-NEXT: vpshufb %ymm6, %ymm10, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512-NEXT: vpshufb %xmm12, %xmm1, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm16
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm14 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-NEXT: vpshufb %xmm14, %xmm2, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm28
-; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512-NEXT: vpshufb %xmm2, %xmm4, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512-NEXT: vpor %xmm0, %xmm3, %xmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-NEXT: vmovdqa (%rdx), %xmm6
; AVX512-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512-NEXT: vpshufb %xmm11, %xmm0, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm31
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512-NEXT: vpshufb %xmm0, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm30
-; AVX512-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa (%r8), %xmm3
-; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[4,u,u,u],zero,xmm1[7],zero,xmm1[5,u,u,u],zero,xmm1[8],zero,xmm1[6]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[4],zero,xmm3[u,u,u,7],zero,xmm3[5],zero,xmm3[u,u,u,8],zero,xmm3[6],zero
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm23
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm5 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512-NEXT: vpshufb %xmm5, %xmm6, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512-NEXT: vpor %xmm0, %xmm3, %xmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%r9), %xmm0
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa (%r8), %xmm7
+; AVX512-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-NEXT: vpshufb %xmm3, %xmm0, %xmm6
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX512-NEXT: vpor %xmm6, %xmm7, %xmm6
+; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX512-NEXT: vpshufb %xmm4, %xmm10, %xmm4
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm13
+; AVX512-NEXT: vpshufb %xmm5, %xmm13, %xmm5
+; AVX512-NEXT: vpor %xmm4, %xmm5, %xmm4
+; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm6
+; AVX512-NEXT: vpshufb %xmm1, %xmm6, %xmm1
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX512-NEXT: vpshufb %xmm2, %xmm9, %xmm2
; AVX512-NEXT: vpor %xmm1, %xmm2, %xmm1
; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero,zero,zero,ymm10[27],zero,ymm10[25]
-; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3
-; 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: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm26, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512-NEXT: vpshufb %ymm13, %ymm10, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm8[25],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512-NEXT: vpshufb %ymm2, %ymm7, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %ymm4, %ymm8, %ymm1
-; AVX512-NEXT: vpshufb %ymm9, %ymm7, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm2
-; AVX512-NEXT: vpshufb %ymm2, %ymm5, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm6[23],zero,ymm6[23,24,25,26],zero,ymm6[24],zero,ymm6[30,31]
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512-NEXT: vpshufb %ymm2, %ymm6, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rax), %ymm4
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm2
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpermi2d %zmm2, %zmm3, %zmm18
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm13
-; AVX512-NEXT: vpshufb %xmm11, %xmm13, %xmm2
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512-NEXT: vpshufb %xmm0, %xmm10, %xmm0
-; AVX512-NEXT: vporq %xmm2, %xmm0, %xmm26
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512-NEXT: vpshufb %xmm12, %xmm11, %xmm0
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm8
-; AVX512-NEXT: vpshufb %xmm14, %xmm8, %xmm2
-; AVX512-NEXT: vporq %xmm0, %xmm2, %xmm24
-; AVX512-NEXT: vmovdqa 32(%rax), %xmm1
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm20 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufb %xmm3, %xmm1, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm25
-; AVX512-NEXT: vpermi2d %zmm0, %zmm2, %zmm20
+; AVX512-NEXT: vmovdqa 32(%rax), %xmm4
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm4[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm4[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm20
; AVX512-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm15[4,u,u,u],zero,xmm15[7],zero,xmm15[5,u,u,u],zero,xmm15[8],zero,xmm15[6]
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm5
-; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm5[4],zero,xmm5[u,u,u,7],zero,xmm5[5],zero,xmm5[u,u,u,8],zero,xmm5[6],zero
-; AVX512-NEXT: vporq %xmm0, %xmm3, %xmm22
-; AVX512-NEXT: vmovdqa (%rax), %xmm2
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
-; 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, %ymm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3],xmm8[4],xmm11[4],xmm8[5],xmm11[5],xmm8[6],xmm11[6],xmm8[7],xmm11[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, %xmm6, %xmm6
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm21
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[2,3,2,3],zmm6[0,1,0,1]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm10[0],xmm13[0],xmm10[1],xmm13[1],xmm10[2],xmm13[2],xmm10[3],xmm13[3],xmm10[4],xmm13[4],xmm10[5],xmm13[5],xmm10[6],xmm13[6],xmm10[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, %xmm7, %xmm7
+; AVX512-NEXT: vpshufb %xmm3, %xmm15, %xmm1
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm11
+; AVX512-NEXT: vpshufb %xmm0, %xmm11, %xmm0
+; AVX512-NEXT: vporq %xmm1, %xmm0, %xmm23
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa64 %ymm21, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm25, %ymm1
+; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm26
+; AVX512-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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 %ymm19, %ymm3
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm3[23],zero,ymm3[21,22,23,26],zero,ymm3[24],zero,ymm3[28,29,26,27]
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm3[18,19,20,21],zero,ymm3[19],zero,ymm3[25,26,27,22],zero,ymm3[20],zero
+; AVX512-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512-NEXT: vmovdqa64 %ymm24, %ymm1
+; AVX512-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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[20],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm1[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-NEXT: vmovdqa64 %ymm17, %ymm1
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm1[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-NEXT: vmovdqa (%rax), %ymm7
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = 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]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm17 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm7[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm17
+; AVX512-NEXT: vmovdqa (%rax), %xmm1
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,5,6]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512-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-NEXT: vpshufb %ymm5, %ymm7, %ymm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3],xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512-NEXT: vmovdqa %xmm9, %xmm12
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm19
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-NEXT: vpshufb %xmm6, %xmm3, %xmm3
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm31[2,3,2,3],zmm3[0,1,0,1]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vmovdqa %xmm10, %xmm9
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-NEXT: vpshufb %xmm8, %xmm4, %xmm4
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm0[2,3,2,3],zmm7[0,1,0,1]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[2,3,2,3],zmm0[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, %ymm17
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm9, %ymm0, %ymm9
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,2]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm9
-; AVX512-NEXT: vmovdqa64 %xmm28, %xmm4
-; AVX512-NEXT: vmovdqa64 %xmm16, %xmm0
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm0[8],xmm4[8],xmm0[9],xmm4[9],xmm0[10],xmm4[10],xmm0[11],xmm4[11],xmm0[12],xmm4[12],xmm0[13],xmm4[13],xmm0[14],xmm4[14],xmm0[15],xmm4[15]
-; AVX512-NEXT: vmovdqa64 %xmm12, %xmm16
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
-; AVX512-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm28 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm28 = zmm4[0,1,0,1],mem[0,1,0,1]
-; AVX512-NEXT: vmovdqa64 %xmm30, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm12
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm0[8],xmm12[9],xmm0[9],xmm12[10],xmm0[10],xmm12[11],xmm0[11],xmm12[12],xmm0[12],xmm12[13],xmm0[13],xmm12[14],xmm0[14],xmm12[15],xmm0[15]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3],xmm0[4],xmm12[4],xmm0[5],xmm12[5],xmm0[6],xmm12[6],xmm0[7],xmm12[7]
-; AVX512-NEXT: vpshufb %xmm14, %xmm12, %xmm12
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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 {{.*#+}} zmm7 = zmm6 ^ (zmm14 & (zmm7 ^ zmm6))
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm6 = mem[2,3,2,3]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm0[2,3,2,3],zmm4[0,1,0,1]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm15[0],xmm11[1],xmm15[1],xmm11[2],xmm15[2],xmm11[3],xmm15[3],xmm11[4],xmm15[4],xmm11[5],xmm15[5],xmm11[6],xmm15[6],xmm11[7],xmm15[7]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm18[2,3,2,3],zmm0[0,1,0,1]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm7 = 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]
+; AVX512-NEXT: vpshufb %ymm5, %ymm14, %ymm5
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[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, %zmm5, %zmm14
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm3[8],xmm0[8],xmm3[9],xmm0[9],xmm3[10],xmm0[10],xmm3[11],xmm0[11],xmm3[12],xmm0[12],xmm3[13],xmm0[13],xmm3[14],xmm0[14],xmm3[15],xmm0[15]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
+; AVX512-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm29 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm29 = zmm0[0,1,0,1],mem[0,1,0,1]
+; AVX512-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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 %xmm8, %xmm6, %xmm6
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm8 = [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 {{.*#+}} zmm16 = zmm31 ^ (zmm8 & (zmm16 ^ zmm31))
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm3 = mem[2,3,2,3]
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm31, %zmm3
; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Folded Reload
; AVX512-NEXT: # ymm31 = mem[2,3,2,3]
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm31
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm31 = zmm6 ^ (zmm14 & (zmm31 ^ zmm6))
-; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm6 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm6 = zmm12[0,1,0,1],mem[0,1,0,1]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm31 = zmm3 ^ (zmm8 & (zmm31 ^ zmm3))
+; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm3 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm3 = zmm6[0,1,0,1],mem[0,1,0,1]
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-NEXT: vmovdqa64 %xmm23, %xmm14
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm14[8],xmm0[8],xmm14[9],xmm0[9],xmm14[10],xmm0[10],xmm14[11],xmm0[11],xmm14[12],xmm0[12],xmm14[13],xmm0[13],xmm14[14],xmm0[14],xmm14[15],xmm0[15]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3],xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512-NEXT: vpshufb %xmm1, %xmm14, %xmm14
-; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm23 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm23 = zmm14[0,1,0,1],mem[0,1,0,1]
-; AVX512-NEXT: vmovdqa64 %xmm27, %xmm0
-; AVX512-NEXT: vpshufb %xmm0, %xmm2, %xmm14
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm0[8],xmm8[8],xmm0[9],xmm8[9],xmm0[10],xmm8[10],xmm0[11],xmm8[11],xmm0[12],xmm8[12],xmm0[13],xmm8[13],xmm0[14],xmm8[14],xmm0[15],xmm8[15]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
+; AVX512-NEXT: vpshufb %xmm10, %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 {{.*#+}} xmm10 = xmm1[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,0,0,4,5,6,7]
; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm13[8],xmm10[8],xmm13[9],xmm10[9],xmm13[10],xmm10[10],xmm13[11],xmm10[11],xmm13[12],xmm10[12],xmm13[13],xmm10[13],xmm13[14],xmm10[14],xmm13[15],xmm10[15]
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm9[8],xmm13[8],xmm9[9],xmm13[9],xmm9[10],xmm13[10],xmm9[11],xmm13[11],xmm9[12],xmm13[12],xmm9[13],xmm13[13],xmm9[14],xmm13[14],xmm9[15],xmm13[15]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm13 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-NEXT: vpshufb %xmm13, %xmm4, %xmm4
+; AVX512-NEXT: vpshufb %xmm13, %xmm4, %xmm0
; AVX512-NEXT: vpshufb %xmm13, %xmm10, %xmm10
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm26[0,1,0,1],zmm10[0,1,0,1]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-NEXT: vmovdqa64 %xmm16, %xmm0
-; AVX512-NEXT: vpshufb %xmm11, %xmm0, %xmm0
-; AVX512-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm24[0,1,0,1],zmm8[0,1,0,1]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-NEXT: vmovdqa {{.*#+}} xmm5 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-NEXT: vpshufb %xmm5, %xmm12, %xmm11
-; AVX512-NEXT: vpshufb %xmm5, %xmm2, %xmm2
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm22[0,1,0,1],zmm2[0,1,0,1]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm12 = zmm12[2,3,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm5))
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm13 = [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 {{.*#+}} zmm5 = zmm5 ^ (zmm13 & (zmm5 ^ zmm12))
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm12 = zmm12[2,3,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm4[0,1,0,1],zmm10[0,1,0,1]
+; AVX512-NEXT: vmovdqa64 %xmm19, %xmm4
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm4[8],xmm12[8],xmm4[9],xmm12[9],xmm4[10],xmm12[10],xmm4[11],xmm12[11],xmm4[12],xmm12[12],xmm4[13],xmm12[13],xmm4[14],xmm12[14],xmm4[15],xmm12[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm13 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-NEXT: vpshufb %xmm13, %xmm5, %xmm5
+; AVX512-NEXT: vpshufb %xmm13, %xmm12, %xmm12
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm4[0,1,0,1],zmm12[0,1,0,1]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm11[8],xmm15[8],xmm11[9],xmm15[9],xmm11[10],xmm15[10],xmm11[11],xmm15[11],xmm11[12],xmm15[12],xmm11[13],xmm15[13],xmm11[14],xmm15[14],xmm11[15],xmm15[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-NEXT: vpshufb %xmm11, %xmm6, %xmm6
+; AVX512-NEXT: vpshufb %xmm11, %xmm9, %xmm9
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm23[0,1,0,1],zmm9[0,1,0,1]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm11 = zmm11[2,3,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm13 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm13 = zmm13[2,3,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm11))
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm11 = zmm11[2,3,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm23 = [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 {{.*#+}} zmm11 = zmm11 ^ (zmm23 & (zmm11 ^ zmm13))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm12 = zmm10 ^ (zmm23 & (zmm12 ^ zmm10))
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm17[2,3,2,3]
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm15
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[1,1,0,0,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm17 = zmm17[2,3,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm17 = zmm12 ^ (zmm13 & (zmm17 ^ zmm12))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm10 ^ (zmm13 & (zmm8 ^ zmm10))
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
+; AVX512-NEXT: vmovdqa64 %xmm28, %xmm4
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm4[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,1,2,0]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm25[2,3,2,3],zmm26[2,3,2,3]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm21[2,3,2,3],zmm22[2,3,2,3]
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm15 = zmm13 ^ (zmm23 & (zmm15 ^ zmm13))
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm0))
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm0 & mem)
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm4))
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm15[0,0,1,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm0
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm19))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm7))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (mem & (zmm29 ^ zmm5))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm5))
+; AVX512-NEXT: vinserti64x4 $1, (%rsp), %zmm6, %zmm5 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm5 & mem)
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm10[0,0,1,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm18))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm16))
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm11))
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512-NEXT: # zmm4 = zmm4[0,1,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm4 & mem)
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm31))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm28))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm4 & mem)
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm31))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm29))
; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,0,4,4,5,4]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm23))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm6))
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
-; AVX512-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (mem & (zmm18 ^ zmm4))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (mem & (zmm18 ^ zmm17))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm20 ^ (mem & (zmm2 ^ zmm20))
-; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm8))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm8))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm20 ^ (mem & (zmm9 ^ zmm20))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm12))
+; AVX512-NEXT: vporq %zmm27, %zmm24, %zmm3
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm3))
+; AVX512-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm15))
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm18, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm17, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm9, 256(%rax)
; AVX512-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm9, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm29, 384(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm5, 384(%rax)
; AVX512-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm3, 64(%rax)
-; AVX512-NEXT: addq $1496, %rsp # imm = 0x5D8
+; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)
+; AVX512-NEXT: addq $1256, %rsp # imm = 0x4E8
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i8_stride7_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $1416, %rsp # imm = 0x588
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm12
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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,zero,zero,zero,zero,ymm12[18]
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm9
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[0,1,14],zero,ymm9[12,13,0,1,14,15],zero,ymm9[3,12,13,2,3,16],zero,ymm9[30,31,28,29,16,17],zero,ymm9[31,18,19,28,29,18],zero
+; AVX512-FCP-NEXT: subq $1448, %rsp # imm = 0x5A8
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm8
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero,zero,zero,ymm8[18]
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %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: vmovdqa64 %ymm2, %ymm16
; 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), %ymm6
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[0,1,0,1,14],zero,ymm4[14,15,0,1,14,15],zero,ymm4[13,14,15,16,17,16],zero,ymm4[30,31,30,31,16,17],zero,ymm4[31,28,29,30,31]
-; AVX512-FCP-NEXT: vpor %ymm1, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm11[13,u,u,u,u,u],zero,ymm11[14,u,u,u,u,u],zero,ymm11[15,u,u,u,u,u],zero,ymm11[16,u,u,u,u,u],zero,ymm11[17,u,u,u]
-; AVX512-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512-FCP-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-FCP-NEXT: vpshufb %ymm2, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512-FCP-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-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm1
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm15, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: vmovdqa (%r8), %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa64 %ymm1, %ymm20
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm9
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[13,u,u,u,u,u],zero,ymm9[14,u,u,u,u,u],zero,ymm9[15,u,u,u,u,u],zero,ymm9[16,u,u,u,u,u],zero,ymm9[17,u,u,u]
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512-FCP-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-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm11, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero,ymm5[27],zero,ymm5[25]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm11[23,u,u,u],zero,ymm11[26],zero,ymm11[24,u,u,u],zero,ymm11[27],zero
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm4
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm7
-; AVX512-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm14
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm1
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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-FCP-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm3
+; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; 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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm4[23],zero,ymm4[21,22,23,26],zero,ymm4[24],zero,ymm4[28,29,26,27]
+; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
; AVX512-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm12
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm2
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-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-FCP-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-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm15, %ymm13
+; AVX512-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[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: vshufi64x2 {{.*#+}} zmm13 = zmm15[2,3,2,3],zmm13[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm8, %ymm21
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm8
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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: # ymm5 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm27
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm14, %ymm19
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm8
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; 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), %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm16
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm7[2,3,2,3],zmm1[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm6
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [4,0,6,0,4,0,6,7,0,17,0,17,0,16,16,0]
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm8[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpermi2d %zmm10, %zmm6, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm7
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm14
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm7
+; AVX512-FCP-NEXT: vmovdqa64 %ymm9, %ymm20
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb {{.*#+}} ymm7 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm9[19],zero,ymm9[21,20,21,22],zero,ymm9[20],zero,ymm9[22,23]
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm7, %xmm6
+; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
+; AVX512-FCP-NEXT: vmovdqa64 %xmm7, %xmm16
+; 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, %xmm0, %xmm7
+; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm27
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm30
+; AVX512-FCP-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm7
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm15, %xmm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm0, %xmm9
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm28
+; AVX512-FCP-NEXT: vpor %xmm6, %xmm9, %xmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-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-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm26
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,4,5,5,6]
+; 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, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm8, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[20],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm9
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm6
+; AVX512-FCP-NEXT: vmovdqa64 %xmm8, %xmm23
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm31
+; 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, %xmm9, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm22
+; AVX512-FCP-NEXT: vpor %xmm6, %xmm8, %xmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm8[18],zero,ymm8[20,21,20,21],zero,ymm8[19],zero,ymm8[19,20,21,22],zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm9, %ymm5
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm5, %ymm30
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm6
; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm2
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm8
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm29 = [4,0,6,0,4,0,6,7,0,17,0,17,0,16,16,0]
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm0, %zmm29
-; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %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[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm4[18,19,20,21],zero,ymm4[19],zero,ymm4[25,26,27,22],zero,ymm4[20],zero
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm11[14],zero,zero,zero,zero,zero,zero,ymm11[15],zero,zero,zero,zero,zero,zero,ymm11[16],zero,zero,zero,zero,zero,zero,ymm11[17],zero,zero,zero,zero,zero,zero,ymm11[18]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[0,1,14],zero,ymm5[12,13,0,1,14,15],zero,ymm5[3,12,13,2,3,16],zero,ymm5[30,31,28,29,16,17],zero,ymm5[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: vmovdqa 32(%rdx), %xmm5
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; 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, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm24
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm5[u,u,u,7],zero,xmm5[5],zero,xmm5[u,u,u,8],zero,xmm5[6],zero,xmm5[u,u]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm23
-; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm12, %xmm5
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm7, %xmm21
-; AVX512-FCP-NEXT: vpor %xmm5, %xmm6, %xmm5
-; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,6]
-; 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, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm16
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm25
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm11
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm5
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm11, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm17
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm5, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm7, %xmm18
-; AVX512-FCP-NEXT: vporq %xmm2, %xmm6, %xmm31
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = 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: vpshufb {{.*#+}} ymm6 = ymm13[0,1,0,1,14],zero,ymm13[14,15,0,1,14,15],zero,ymm13[13,14,15,16,17,16],zero,ymm13[30,31,30,31,16,17],zero,ymm13[31,28,29,30,31]
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm6, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm3
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm6
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero,zero,zero,ymm6[18]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[0,1,14],zero,ymm15[12,13,0,1,14,15],zero,ymm15[3,12,13,2,3,16],zero,ymm15[30,31,28,29,16,17],zero,ymm15[31,18,19,28,29,18],zero
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20],zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm15[19],zero,ymm15[21,20,21,22],zero,ymm15[20],zero,ymm15[22,23]
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[20],zero,ymm10[18],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm14[18],zero,ymm14[20,21,20,21],zero,ymm14[19],zero,ymm14[19,20,21,22],zero
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,ymm14[14],zero,zero,zero,zero,zero,zero,ymm14[15],zero,zero,zero,zero,zero,zero,ymm14[16],zero,zero,zero,zero,zero,zero,ymm14[17],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[13,u,u,u,u,u],zero,ymm10[14,u,u,u,u,u],zero,ymm10[15,u,u,u,u,u],zero,ymm10[16,u,u,u,u,u],zero,ymm10[17,u,u,u]
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm10
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm1
+; 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, %ymm5, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
+; 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,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm3[18],zero,ymm3[20,21,20,21],zero,ymm3[19],zero,ymm3[19,20,21,22],zero
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,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
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = 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 %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm1
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm2
+; AVX512-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm11, %xmm1
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm6
+; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm6, %xmm2
+; AVX512-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm1
; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512-FCP-NEXT: vporq %xmm1, %xmm0, %xmm20
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm13, %xmm0
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm8[u,u,u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u]
-; AVX512-FCP-NEXT: vporq %xmm0, %xmm1, %xmm19
-; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm1
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,0,1,0,0,0,0,16,0,16,0,18,19,0,17]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm16
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,0,1,0,0,0,0,16,0,16,0,18,19,0,17]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm10[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm18
; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm7
-; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm23, %xmm1
; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm7, %xmm3
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm6, %xmm15
-; AVX512-FCP-NEXT: vporq %xmm3, %xmm15, %xmm18
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm5
+; AVX512-FCP-NEXT: vmovdqa64 %xmm22, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm4
+; AVX512-FCP-NEXT: vporq %xmm3, %xmm4, %xmm26
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = 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-FCP-NEXT: vmovdqa64 %ymm21, %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[23,u,u,u],zero,ymm1[26],zero,ymm1[24,u,u,u],zero,ymm1[27],zero
+; AVX512-FCP-NEXT: vporq %ymm3, %ymm4, %ymm25
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm24
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %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,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm4
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm4[23],zero,ymm4[21,22,23,26],zero,ymm4[24],zero,ymm4[28,29,26,27]
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm23
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm4[18,19,20,21],zero,ymm4[19],zero,ymm4[25,26,27,22],zero,ymm4[20],zero
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm22
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[20],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22]
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm3[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm14[23],zero,ymm14[23,24,25,26],zero,ymm14[24],zero,ymm14[30,31]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm14[18],zero,ymm14[20,21,20,21],zero,ymm14[19],zero,ymm14[19,20,21,22],zero
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm3[2,3,2,3],zmm0[2,3,2,3]
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,5,5,6]
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [2,2,3,3,2,2,3,3]
; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0
; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [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 %ymm1, %ymm4, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm29
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm14
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = 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-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [4,5,4,5,5,7,4,5,20,21,22,23,20,21,22,23]
-; AVX512-FCP-NEXT: vpermt2d %zmm15, %zmm17, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm3
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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]
-; 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, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm27
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm30[2,3,2,3],zmm4[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm23, %xmm1
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm15
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm22
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[2,3,2,3],zmm4[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3],xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm26
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm2[2,3,2,3],zmm4[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm1[8],xmm15[9],xmm1[9],xmm15[10],xmm1[10],xmm15[11],xmm1[11],xmm15[12],xmm1[12],xmm15[13],xmm1[13],xmm15[14],xmm1[14],xmm15[15],xmm1[15]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm13[8],xmm8[8],xmm13[9],xmm8[9],xmm13[10],xmm8[10],xmm13[11],xmm8[11],xmm13[12],xmm8[12],xmm13[13],xmm8[13],xmm13[14],xmm8[14],xmm13[15],xmm8[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm15
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm12[8],xmm3[8],xmm12[9],xmm3[9],xmm12[10],xmm3[10],xmm12[11],xmm3[11],xmm12[12],xmm3[12],xmm12[13],xmm3[13],xmm12[14],xmm3[14],xmm12[15],xmm3[15]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm12, %xmm12
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm10
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3],xmm4[4],xmm10[4],xmm4[5],xmm10[5],xmm4[6],xmm10[6],xmm4[7],xmm10[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm1
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3],xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm10[8],xmm4[8],xmm10[9],xmm4[9],xmm10[10],xmm4[10],xmm10[11],xmm4[11],xmm10[12],xmm4[12],xmm10[13],xmm4[13],xmm10[14],xmm4[14],xmm10[15],xmm4[15]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm10
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm3[0,1,0,1],zmm0[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm15[8],xmm2[8],xmm15[9],xmm2[9],xmm15[10],xmm2[10],xmm15[11],xmm2[11],xmm15[12],xmm2[12],xmm15[13],xmm2[13],xmm15[14],xmm2[14],xmm15[15],xmm2[15]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm0[0,1,0,1],zmm3[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8],xmm1[8],xmm9[9],xmm1[9],xmm9[10],xmm1[10],xmm9[11],xmm1[11],xmm9[12],xmm1[12],xmm9[13],xmm1[13],xmm9[14],xmm1[14],xmm9[15],xmm1[15]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm0
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm11[8],xmm5[9],xmm11[9],xmm5[10],xmm11[10],xmm5[11],xmm11[11],xmm5[12],xmm11[12],xmm5[13],xmm11[13],xmm5[14],xmm11[14],xmm5[15],xmm11[15]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm11
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm1
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm31[0,1,0,1],zmm1[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpermd %ymm2, %ymm17, %ymm2
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm20[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3],xmm8[4],xmm13[4],xmm8[5],xmm13[5],xmm8[6],xmm13[6],xmm8[7],xmm13[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm22, %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm8
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm19[0,1,0,1]
-; 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: vmovdqa64 %xmm26, %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,0,1],zmm18[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm7 = zmm7[2,3,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm9 = zmm9[2,3,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [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 {{.*#+}} zmm9 = zmm7 ^ (zmm10 & (zmm9 ^ zmm7))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm7))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm9))
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm2[0,1,0,1],zmm0[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshuflw {{.*#+}} ymm15 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,5,4,5,5,7,4,5,20,21,22,23,20,21,22,23]
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm15
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpermd %ymm13, %ymm28, %ymm13
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3],xmm6[4],xmm11[4],xmm6[5],xmm11[5],xmm6[6],xmm11[6],xmm6[7],xmm11[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm6 = zmm6[0,1,0,1],mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm26[0,1,0,1]
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm24[2,3,2,3],zmm25[2,3,2,3]
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm22[2,3,2,3],zmm23[2,3,2,3]
+; AVX512-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm5 ^ (zmm12 & (zmm7 ^ zmm5))
+; AVX512-FCP-NEXT: vporq %zmm20, %zmm19, %zmm5
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm5))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm7))
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm0 = zmm0[2,3,2,3],mem[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
; AVX512-FCP-NEXT: # zmm7 = zmm7[2,3,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm12[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm0))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm0 = zmm0[2,3,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm10 & (zmm0 ^ zmm7))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm24 ^ (zmm10 & (zmm21 ^ zmm24))
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm12 & (zmm5 ^ zmm7))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm31 = zmm17 ^ (zmm12 & (zmm31 ^ zmm17))
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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 {{.*#+}} zmm23 = zmm30 ^ (zmm7 & (zmm23 ^ zmm30))
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm10 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm21 ^ (zmm7 & (zmm16 ^ zmm21))
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm11 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm12 = mem[2,3,2,3]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm13, %zmm10
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm13 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm11 ^ (zmm7 & (zmm12 ^ zmm11))
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm7 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm4))
+; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm3 & mem)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm7))
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm13
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm10 ^ (zmm7 & (zmm13 ^ zmm10))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm12, %zmm9 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm7))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm7 & mem)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm9))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm0))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm29 ^ (mem & (zmm4 ^ zmm29))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm23))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm25 ^ (mem & (zmm5 ^ zmm25))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm21))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & mem)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm13))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm2))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm16 ^ (mem & (zmm6 ^ zmm16))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm4 ^ (mem & (zmm30 ^ zmm4))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (mem & (zmm30 ^ zmm16))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm1 ^ (mem & (zmm9 ^ zmm1))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm31))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm1 = zmm1[0,1,2,3],mem[2,3,2,3]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 | (zmm1 & mem)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm12))
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm8))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm18 ^ (mem & (zmm2 ^ zmm18))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm6))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 256(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 384(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 64(%rax)
-; AVX512-FCP-NEXT: addq $1416, %rsp # imm = 0x588
+; AVX512-FCP-NEXT: addq $1448, %rsp # imm = 0x5A8
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i8_stride7_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $1496, %rsp # imm = 0x5D8
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [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 %ymm1, %ymm3, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-NEXT: subq $1256, %rsp # imm = 0x4E8
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm5 = [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 %ymm5, %ymm9, %ymm0
; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [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 %ymm2, %ymm10, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm4 = [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 %ymm4, %ymm10, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm8
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm11
; 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, %ymm8, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm7
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm7, %ymm1
-; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm6
-; AVX512DQ-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-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm24
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm5
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm13
+; 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, %ymm13, %ymm2
+; AVX512DQ-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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: # ymm14 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm3, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [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 %ymm0, %ymm10, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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: # ymm15 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm7, %ymm1
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm8, %ymm3
-; AVX512DQ-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm6, %ymm1
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: vmovdqa (%r8), %ymm14
+; 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, %ymm14, %ymm0
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm15
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm15, %ymm7
+; AVX512DQ-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm9, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm9, %ymm21
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm10, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm10, %ymm18
+; AVX512DQ-NEXT: vporq %ymm0, %ymm7, %ymm31
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm5, %ymm13
-; AVX512DQ-NEXT: vpor %ymm1, %ymm13, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm13, %ymm1
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm14
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm14, %ymm0
-; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm13, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm13, %ymm19
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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: # ymm10 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm11, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm11, %ymm16
+; AVX512DQ-NEXT: vpor %ymm0, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero,zero,zero,ymm14[27],zero,ymm14[25]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm13[23,u,u,u],zero,ymm13[26],zero,ymm13[24,u,u,u],zero,ymm13[27],zero
-; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; 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, %ymm14, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm14, %ymm17
+; 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, %ymm15, %ymm13
+; AVX512DQ-NEXT: vmovdqa64 %ymm15, %ymm24
+; AVX512DQ-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm7
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm7, %ymm13
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm8
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm8, %ymm12
+; AVX512DQ-NEXT: vpor %ymm13, %ymm12, %ymm12
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero,zero,zero,ymm8[27],zero,ymm8[25]
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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-NEXT: # ymm14 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm14, %ymm7, %ymm13
+; AVX512DQ-NEXT: vmovdqa64 %ymm14, %ymm25
+; AVX512DQ-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm12, %ymm9
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm13, %ymm10
+; AVX512DQ-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm13[25],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm12[23],zero,ymm12[21,22,23,26],zero,ymm12[24],zero,ymm12[28,29,26,27]
+; AVX512DQ-NEXT: vpor %ymm9, %ymm10, %ymm9
+; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm9
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm9, %ymm0
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm10, %ymm11
+; AVX512DQ-NEXT: vpor %ymm0, %ymm11, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm10[25],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm11 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm9[23],zero,ymm9[23,24,25,26],zero,ymm9[24],zero,ymm9[30,31]
+; AVX512DQ-NEXT: vpor %ymm0, %ymm11, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm1, %ymm15
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm0, %ymm2
-; AVX512DQ-NEXT: vpor %ymm2, %ymm15, %ymm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm0[25],zero,ymm0[23],zero,zero,zero,zero,ymm0[26],zero,ymm0[24],zero,zero,zero,zero
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm1, %ymm15
-; AVX512DQ-NEXT: vmovdqa64 %ymm12, %ymm22
-; AVX512DQ-NEXT: vpor %ymm2, %ymm15, %ymm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm15
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm15, %ymm3
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm2, %ymm9
-; AVX512DQ-NEXT: vpor %ymm3, %ymm9, %ymm3
-; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm2, %ymm3
-; AVX512DQ-NEXT: vmovdqa64 %ymm9, %ymm21
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm9 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm15[23],zero,ymm15[23,24,25,26],zero,ymm15[24],zero,ymm15[30,31]
-; AVX512DQ-NEXT: vpor %ymm3, %ymm9, %ymm3
-; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm9[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 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm9, %ymm12
-; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm20
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm12[2,3,2,3],zmm3[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm0, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512DQ-NEXT: vpor %ymm3, %ymm11, %ymm3
-; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm0, %ymm0
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm1, %ymm1
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm14
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[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 {{.*#+}} ymm11 = 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]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm11[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm12, %ymm1
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[18],zero,zero,zero,zero,ymm13[21],zero,ymm13[19],zero,zero,zero,zero,ymm13[22],zero,ymm13[20]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm12[18,19,20,21],zero,ymm12[19],zero,ymm12[25,26,27,22],zero,ymm12[20],zero
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm7, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm8, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm26
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm14, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm7, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm12 = [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 %ymm12, %ymm8, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm15, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[20],zero,ymm10[18],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22]
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm8, %ymm9, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm15, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm9, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm10, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm12 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm1, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm16
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm14 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm2, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm28
-; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm4, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512DQ-NEXT: vpor %xmm0, %xmm3, %xmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm6
; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm11, %xmm0, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm31
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm30
-; AVX512DQ-NEXT: vpor %xmm1, %xmm2, %xmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm3
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[4,u,u,u],zero,xmm1[7],zero,xmm1[5,u,u,u],zero,xmm1[8],zero,xmm1[6]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[4],zero,xmm3[u,u,u,7],zero,xmm3[5],zero,xmm3[u,u,u,8],zero,xmm3[6],zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm23
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm5 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm5, %xmm6, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512DQ-NEXT: vpor %xmm0, %xmm3, %xmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm0
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm7
+; AVX512DQ-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm0, %xmm6
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm7, %xmm7
+; AVX512DQ-NEXT: vpor %xmm6, %xmm7, %xmm6
+; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm10, %xmm4
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm13
+; AVX512DQ-NEXT: vpshufb %xmm5, %xmm13, %xmm5
+; AVX512DQ-NEXT: vpor %xmm4, %xmm5, %xmm4
+; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm6
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm6, %xmm1
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm9, %xmm2
; AVX512DQ-NEXT: vpor %xmm1, %xmm2, %xmm1
; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero,zero,zero,ymm10[27],zero,ymm10[25]
-; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3
-; 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: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm10, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm8[25],zero,ymm8[23],zero,zero,zero,zero,ymm8[26],zero,ymm8[24],zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm7, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm8, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm7, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm6[23],zero,ymm6[23,24,25,26],zero,ymm6[24],zero,ymm6[30,31]
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm6, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm2
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm3, %zmm18
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm13
-; AVX512DQ-NEXT: vpshufb %xmm11, %xmm13, %xmm2
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm10, %xmm0
-; AVX512DQ-NEXT: vporq %xmm2, %xmm0, %xmm26
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm11, %xmm0
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm8
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm8, %xmm2
-; AVX512DQ-NEXT: vporq %xmm0, %xmm2, %xmm24
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm1
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm20 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufb %xmm3, %xmm1, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm25
-; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm2, %zmm20
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm4
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm4[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm4[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm20
; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm15[4,u,u,u],zero,xmm15[7],zero,xmm15[5,u,u,u],zero,xmm15[8],zero,xmm15[6]
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm5
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm5[4],zero,xmm5[u,u,u,7],zero,xmm5[5],zero,xmm5[u,u,u,8],zero,xmm5[6],zero
-; AVX512DQ-NEXT: vporq %xmm0, %xmm3, %xmm22
-; AVX512DQ-NEXT: vmovdqa (%rax), %xmm2
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
-; 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, %ymm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3],xmm8[4],xmm11[4],xmm8[5],xmm11[5],xmm8[6],xmm11[6],xmm8[7],xmm11[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, %xmm6, %xmm6
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm21
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[2,3,2,3],zmm6[0,1,0,1]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm10[0],xmm13[0],xmm10[1],xmm13[1],xmm10[2],xmm13[2],xmm10[3],xmm13[3],xmm10[4],xmm13[4],xmm10[5],xmm13[5],xmm10[6],xmm13[6],xmm10[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, %xmm7, %xmm7
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm15, %xmm1
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm11
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm11, %xmm0
+; AVX512DQ-NEXT: vporq %xmm1, %xmm0, %xmm23
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa64 %ymm21, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm26
+; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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 %ymm19, %ymm3
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm3[23],zero,ymm3[21,22,23,26],zero,ymm3[24],zero,ymm3[28,29,26,27]
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm3[18,19,20,21],zero,ymm3[19],zero,ymm3[25,26,27,22],zero,ymm3[20],zero
+; AVX512DQ-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512DQ-NEXT: vmovdqa64 %ymm24, %ymm1
+; AVX512DQ-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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[20],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm1[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm1
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm1[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqa (%rax), %ymm7
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = 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]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm17 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm7[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm17
+; AVX512DQ-NEXT: vmovdqa (%rax), %xmm1
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,5,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512DQ-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-NEXT: vpshufb %ymm5, %ymm7, %ymm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3],xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512DQ-NEXT: vmovdqa %xmm9, %xmm12
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm19
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm3, %xmm3
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm31[2,3,2,3],zmm3[0,1,0,1]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vmovdqa %xmm10, %xmm9
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-NEXT: vpshufb %xmm8, %xmm4, %xmm4
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm0[2,3,2,3],zmm7[0,1,0,1]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[2,3,2,3],zmm0[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, %ymm17
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm0, %ymm9
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,2]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm9
-; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm16, %xmm0
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm0[8],xmm4[8],xmm0[9],xmm4[9],xmm0[10],xmm4[10],xmm0[11],xmm4[11],xmm0[12],xmm4[12],xmm0[13],xmm4[13],xmm0[14],xmm4[14],xmm0[15],xmm4[15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm12, %xmm16
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm28 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm28 = zmm4[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm12
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm0[8],xmm12[9],xmm0[9],xmm12[10],xmm0[10],xmm12[11],xmm0[11],xmm12[12],xmm0[12],xmm12[13],xmm0[13],xmm12[14],xmm0[14],xmm12[15],xmm0[15]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3],xmm0[4],xmm12[4],xmm0[5],xmm12[5],xmm0[6],xmm12[6],xmm0[7],xmm12[7]
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm12, %xmm12
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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 {{.*#+}} zmm7 = zmm6 ^ (zmm14 & (zmm7 ^ zmm6))
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm6 = mem[2,3,2,3]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm0[2,3,2,3],zmm4[0,1,0,1]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm15[0],xmm11[1],xmm15[1],xmm11[2],xmm15[2],xmm11[3],xmm15[3],xmm11[4],xmm15[4],xmm11[5],xmm15[5],xmm11[6],xmm15[6],xmm11[7],xmm15[7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm10 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm18[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm7 = 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]
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm14, %ymm5
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[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, %zmm5, %zmm14
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm3[8],xmm0[8],xmm3[9],xmm0[9],xmm3[10],xmm0[10],xmm3[11],xmm0[11],xmm3[12],xmm0[12],xmm3[13],xmm0[13],xmm3[14],xmm0[14],xmm3[15],xmm0[15]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm29 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm29 = zmm0[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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 %xmm8, %xmm6, %xmm6
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm8 = [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 {{.*#+}} zmm16 = zmm31 ^ (zmm8 & (zmm16 ^ zmm31))
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm3 = mem[2,3,2,3]
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm31, %zmm3
; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Folded Reload
; AVX512DQ-NEXT: # ymm31 = mem[2,3,2,3]
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm31
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm31 = zmm6 ^ (zmm14 & (zmm31 ^ zmm6))
-; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm6 = zmm12[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm31 = zmm3 ^ (zmm8 & (zmm31 ^ zmm3))
+; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm3 = zmm6[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vmovdqa64 %xmm23, %xmm14
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm14[8],xmm0[8],xmm14[9],xmm0[9],xmm14[10],xmm0[10],xmm14[11],xmm0[11],xmm14[12],xmm0[12],xmm14[13],xmm0[13],xmm14[14],xmm0[14],xmm14[15],xmm0[15]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3],xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm14, %xmm14
-; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm23 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm23 = zmm14[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm2, %xmm14
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm0[8],xmm8[8],xmm0[9],xmm8[9],xmm0[10],xmm8[10],xmm0[11],xmm8[11],xmm0[12],xmm8[12],xmm0[13],xmm8[13],xmm0[14],xmm8[14],xmm0[15],xmm8[15]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
+; AVX512DQ-NEXT: vpshufb %xmm10, %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 {{.*#+}} xmm10 = xmm1[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,1,0,0,4,5,6,7]
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm13[8],xmm10[8],xmm13[9],xmm10[9],xmm13[10],xmm10[10],xmm13[11],xmm10[11],xmm13[12],xmm10[12],xmm13[13],xmm10[13],xmm13[14],xmm10[14],xmm13[15],xmm10[15]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm9[8],xmm13[8],xmm9[9],xmm13[9],xmm9[10],xmm13[10],xmm9[11],xmm13[11],xmm9[12],xmm13[12],xmm9[13],xmm13[13],xmm9[14],xmm13[14],xmm9[15],xmm13[15]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm13 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-NEXT: vpshufb %xmm13, %xmm4, %xmm4
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm4, %xmm0
; AVX512DQ-NEXT: vpshufb %xmm13, %xmm10, %xmm10
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm26[0,1,0,1],zmm10[0,1,0,1]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-NEXT: vmovdqa64 %xmm16, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm11, %xmm0, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm24[0,1,0,1],zmm8[0,1,0,1]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-NEXT: vmovdqa {{.*#+}} xmm5 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-NEXT: vpshufb %xmm5, %xmm12, %xmm11
-; AVX512DQ-NEXT: vpshufb %xmm5, %xmm2, %xmm2
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm22[0,1,0,1],zmm2[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm12 = zmm12[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm5))
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm13 = [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 {{.*#+}} zmm5 = zmm5 ^ (zmm13 & (zmm5 ^ zmm12))
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm12 = zmm12[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm4[0,1,0,1],zmm10[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa64 %xmm19, %xmm4
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm4[8],xmm12[8],xmm4[9],xmm12[9],xmm4[10],xmm12[10],xmm4[11],xmm12[11],xmm4[12],xmm12[12],xmm4[13],xmm12[13],xmm4[14],xmm12[14],xmm4[15],xmm12[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm13 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm5, %xmm5
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm12, %xmm12
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm4[0,1,0,1],zmm12[0,1,0,1]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm11[8],xmm15[8],xmm11[9],xmm15[9],xmm11[10],xmm15[10],xmm11[11],xmm15[11],xmm11[12],xmm15[12],xmm11[13],xmm15[13],xmm11[14],xmm15[14],xmm11[15],xmm15[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm6, %xmm6
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm9, %xmm9
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm23[0,1,0,1],zmm9[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm11 = zmm11[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm13 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm13 = zmm13[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm11))
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm11 = zmm11[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm23 = [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 {{.*#+}} zmm11 = zmm11 ^ (zmm23 & (zmm11 ^ zmm13))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm12 = zmm10 ^ (zmm23 & (zmm12 ^ zmm10))
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm17[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm15
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[1,1,0,0,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm17 = zmm17[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm17 = zmm12 ^ (zmm13 & (zmm17 ^ zmm12))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm10 ^ (zmm13 & (zmm8 ^ zmm10))
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm4
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm10 = xmm4[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,1,2,0]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm25[2,3,2,3],zmm26[2,3,2,3]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm21[2,3,2,3],zmm22[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm15 = zmm13 ^ (zmm23 & (zmm15 ^ zmm13))
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm0))
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm0 & mem)
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm4))
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm15[0,0,1,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm0
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm19))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm7))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (mem & (zmm29 ^ zmm5))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm5))
+; AVX512DQ-NEXT: vinserti64x4 $1, (%rsp), %zmm6, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | (zmm5 & mem)
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm10[0,0,1,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm18))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm16))
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm11))
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
; AVX512DQ-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
; AVX512DQ-NEXT: # zmm4 = zmm4[0,1,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | (zmm4 & mem)
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm31))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm28))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm4 & mem)
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm31))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm29))
; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,1,0,4,4,5,4]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm23))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm6))
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (mem & (zmm18 ^ zmm4))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (mem & (zmm18 ^ zmm17))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm20 ^ (mem & (zmm2 ^ zmm20))
-; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm8))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm8))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm20 ^ (mem & (zmm9 ^ zmm20))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm12))
+; AVX512DQ-NEXT: vporq %zmm27, %zmm24, %zmm3
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm3))
+; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (mem & (zmm17 ^ zmm15))
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm18, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm17, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm9, 256(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm9, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm29, 384(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm5, 384(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm3, 64(%rax)
-; AVX512DQ-NEXT: addq $1496, %rsp # imm = 0x5D8
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 64(%rax)
+; AVX512DQ-NEXT: addq $1256, %rsp # imm = 0x4E8
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i8_stride7_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $1416, %rsp # imm = 0x588
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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,zero,zero,zero,zero,ymm12[18]
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[0,1,14],zero,ymm9[12,13,0,1,14,15],zero,ymm9[3,12,13,2,3,16],zero,ymm9[30,31,28,29,16,17],zero,ymm9[31,18,19,28,29,18],zero
+; AVX512DQ-FCP-NEXT: subq $1448, %rsp # imm = 0x5A8
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero,zero,zero,ymm8[18]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %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: vmovdqa64 %ymm2, %ymm16
; 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), %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[0,1,0,1,14],zero,ymm4[14,15,0,1,14,15],zero,ymm4[13,14,15,16,17,16],zero,ymm4[30,31,30,31,16,17],zero,ymm4[31,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm11[13,u,u,u,u,u],zero,ymm11[14,u,u,u,u,u],zero,ymm11[15,u,u,u,u,u],zero,ymm11[16,u,u,u,u,u],zero,ymm11[17,u,u,u]
-; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm2, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm15, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: vmovdqa (%r8), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa64 %ymm1, %ymm20
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[13,u,u,u,u,u],zero,ymm9[14,u,u,u,u,u],zero,ymm9[15,u,u,u,u,u],zero,ymm9[16,u,u,u,u,u],zero,ymm9[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512DQ-FCP-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-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm5, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[23],zero,zero,zero,zero,ymm5[26],zero,ymm5[24],zero,zero,zero,zero,ymm5[27],zero,ymm5[25]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm11[23,u,u,u],zero,ymm11[26],zero,ymm11[24,u,u,u],zero,ymm11[27],zero
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm4
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm7
-; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm14
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm1
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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-FCP-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm3
+; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; 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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm4[23],zero,ymm4[21,22,23,26],zero,ymm4[24],zero,ymm4[28,29,26,27]
+; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
; AVX512DQ-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm2
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm10, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512DQ-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vpor %ymm12, %ymm13, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[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: vshufi64x2 {{.*#+}} zmm13 = zmm15[2,3,2,3],zmm13[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm8, %ymm21
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm8
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm27
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm14, %ymm19
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; 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), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm16
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm7[2,3,2,3],zmm1[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm7
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm6
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [4,0,6,0,4,0,6,7,0,17,0,17,0,16,16,0]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm8
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm8[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm10, %zmm6, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb {{.*#+}} ymm7 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm9[19],zero,ymm9[21,20,21,22],zero,ymm9[20],zero,ymm9[22,23]
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm9, %ymm20
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm7
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm7, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm7, %xmm16
+; 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, %xmm0, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm27
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm30
+; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-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-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm26
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm15, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm0, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm28
+; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm9, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[20],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,4,5,5,6]
+; 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, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm8, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm8[18],zero,ymm8[20,21,20,21],zero,ymm8[19],zero,ymm8[19,20,21,22],zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm8, %xmm23
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm31
+; 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, %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm22
+; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm8, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm9, %ymm5
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm5, %ymm30
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm6
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm2
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm8
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm8, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm29 = [4,0,6,0,4,0,6,7,0,17,0,17,0,16,16,0]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm2
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm0, %zmm29
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %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[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm4[18,19,20,21],zero,ymm4[19],zero,ymm4[25,26,27,22],zero,ymm4[20],zero
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm11[14],zero,zero,zero,zero,zero,zero,ymm11[15],zero,zero,zero,zero,zero,zero,ymm11[16],zero,zero,zero,zero,zero,zero,ymm11[17],zero,zero,zero,zero,zero,zero,ymm11[18]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[0,1,14],zero,ymm5[12,13,0,1,14,15],zero,ymm5[3,12,13,2,3,16],zero,ymm5[30,31,28,29,16,17],zero,ymm5[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: vmovdqa 32(%rdx), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; 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, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm24
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm5[u,u,u,7],zero,xmm5[5],zero,xmm5[u,u,u,8],zero,xmm5[6],zero,xmm5[u,u]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm23
-; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm12, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm7, %xmm21
-; AVX512DQ-FCP-NEXT: vpor %xmm5, %xmm6, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,6]
-; 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, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm16
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm25
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm11, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm17
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm5, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm7, %xmm18
-; AVX512DQ-FCP-NEXT: vporq %xmm2, %xmm6, %xmm31
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = 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: vpshufb {{.*#+}} ymm6 = ymm13[0,1,0,1,14],zero,ymm13[14,15,0,1,14,15],zero,ymm13[13,14,15,16,17,16],zero,ymm13[30,31,30,31,16,17],zero,ymm13[31,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm6, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm3
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,ymm6[14],zero,zero,zero,zero,zero,zero,ymm6[15],zero,zero,zero,zero,zero,zero,ymm6[16],zero,zero,zero,zero,zero,zero,ymm6[17],zero,zero,zero,zero,zero,zero,ymm6[18]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[0,1,14],zero,ymm15[12,13,0,1,14,15],zero,ymm15[3,12,13,2,3,16],zero,ymm15[30,31,28,29,16,17],zero,ymm15[31,18,19,28,29,18],zero
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20],zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21],zero,ymm15[19],zero,ymm15[21,20,21,22],zero,ymm15[20],zero,ymm15[22,23]
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm10[20],zero,ymm10[18],zero,zero,zero,zero,ymm10[21],zero,ymm10[19],zero,zero,zero,zero,ymm10[22]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm14[18],zero,ymm14[20,21,20,21],zero,ymm14[19],zero,ymm14[19,20,21,22],zero
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,ymm14[14],zero,zero,zero,zero,zero,zero,ymm14[15],zero,zero,zero,zero,zero,zero,ymm14[16],zero,zero,zero,zero,zero,zero,ymm14[17],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[13,u,u,u,u,u],zero,ymm10[14,u,u,u,u,u],zero,ymm10[15,u,u,u,u,u],zero,ymm10[16,u,u,u,u,u],zero,ymm10[17,u,u,u]
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm10
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm1
+; 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, %ymm5, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
+; 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,ymm2[20],zero,ymm2[18],zero,zero,zero,zero,ymm2[21],zero,ymm2[19],zero,zero,zero,zero,ymm2[22]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm3[18],zero,ymm3[20,21,20,21],zero,ymm3[19],zero,ymm3[19,20,21,22],zero
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm4, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,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
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = 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 %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm2
+; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm11, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm6, %xmm2
+; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm1
; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-FCP-NEXT: vporq %xmm1, %xmm0, %xmm20
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm13, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm8[u,u,u,7],zero,xmm8[5],zero,xmm8[u,u,u,8],zero,xmm8[6],zero,xmm8[u,u]
-; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm1, %xmm19
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm1
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,0,1,0,0,0,0,16,0,16,0,18,19,0,17]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm16
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,0,1,0,0,0,0,16,0,16,0,18,19,0,17]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm10[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm18
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm23, %xmm1
; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm7, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm6, %xmm15
-; AVX512DQ-FCP-NEXT: vporq %xmm3, %xmm15, %xmm18
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,6]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm22, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT: vporq %xmm3, %xmm4, %xmm26
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = 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-FCP-NEXT: vmovdqa64 %ymm21, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm1[23,u,u,u],zero,ymm1[26],zero,ymm1[24,u,u,u],zero,ymm1[27],zero
+; AVX512DQ-FCP-NEXT: vporq %ymm3, %ymm4, %ymm25
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm24
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %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,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm4[23],zero,ymm4[21,22,23,26],zero,ymm4[24],zero,ymm4[28,29,26,27]
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm4[18,19,20,21],zero,ymm4[19],zero,ymm4[25,26,27,22],zero,ymm4[20],zero
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm22
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %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[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[20],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22]
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm3[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm14[23],zero,ymm14[23,24,25,26],zero,ymm14[24],zero,ymm14[30,31]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm14[18],zero,ymm14[20,21,20,21],zero,ymm14[19],zero,ymm14[19,20,21,22],zero
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm3[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,5,5,6]
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [2,2,3,3,2,2,3,3]
; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0
; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [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 %ymm1, %ymm4, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm29
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm14
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = 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-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm17 = [4,5,4,5,5,7,4,5,20,21,22,23,20,21,22,23]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm15, %zmm17, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm3
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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-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, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm27
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm30[2,3,2,3],zmm4[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm23, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm15
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm22
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[2,3,2,3],zmm4[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3],xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm26
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm2[2,3,2,3],zmm4[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm1[8],xmm15[9],xmm1[9],xmm15[10],xmm1[10],xmm15[11],xmm1[11],xmm15[12],xmm1[12],xmm15[13],xmm1[13],xmm15[14],xmm1[14],xmm15[15],xmm1[15]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm13[8],xmm8[8],xmm13[9],xmm8[9],xmm13[10],xmm8[10],xmm13[11],xmm8[11],xmm13[12],xmm8[12],xmm13[13],xmm8[13],xmm13[14],xmm8[14],xmm13[15],xmm8[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm15
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm12[8],xmm3[8],xmm12[9],xmm3[9],xmm12[10],xmm3[10],xmm12[11],xmm3[11],xmm12[12],xmm3[12],xmm12[13],xmm3[13],xmm12[14],xmm3[14],xmm12[15],xmm3[15]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm12, %xmm12
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm10
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm10[0],xmm4[1],xmm10[1],xmm4[2],xmm10[2],xmm4[3],xmm10[3],xmm4[4],xmm10[4],xmm4[5],xmm10[5],xmm4[6],xmm10[6],xmm4[7],xmm10[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm1
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3],xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm10[8],xmm4[8],xmm10[9],xmm4[9],xmm10[10],xmm4[10],xmm10[11],xmm4[11],xmm10[12],xmm4[12],xmm10[13],xmm4[13],xmm10[14],xmm4[14],xmm10[15],xmm4[15]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm10
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm3[0,1,0,1],zmm0[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm15[8],xmm2[8],xmm15[9],xmm2[9],xmm15[10],xmm2[10],xmm15[11],xmm2[11],xmm15[12],xmm2[12],xmm15[13],xmm2[13],xmm15[14],xmm2[14],xmm15[15],xmm2[15]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm0[0,1,0,1],zmm3[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8],xmm1[8],xmm9[9],xmm1[9],xmm9[10],xmm1[10],xmm9[11],xmm1[11],xmm9[12],xmm1[12],xmm9[13],xmm1[13],xmm9[14],xmm1[14],xmm9[15],xmm1[15]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm11[8],xmm5[9],xmm11[9],xmm5[10],xmm11[10],xmm5[11],xmm11[11],xmm5[12],xmm11[12],xmm5[13],xmm11[13],xmm5[14],xmm11[14],xmm5[15],xmm11[15]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm1
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm31[0,1,0,1],zmm1[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm17, %ymm2
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm20[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm13[0],xmm8[1],xmm13[1],xmm8[2],xmm13[2],xmm8[3],xmm13[3],xmm8[4],xmm13[4],xmm8[5],xmm13[5],xmm8[6],xmm13[6],xmm8[7],xmm13[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm22, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm8
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm19[0,1,0,1]
-; 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: vmovdqa64 %xmm26, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,0,1],zmm18[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm7 = zmm7[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm9 = zmm9[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [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 {{.*#+}} zmm9 = zmm7 ^ (zmm10 & (zmm9 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm2[0,1,0,1],zmm0[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshuflw {{.*#+}} ymm15 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,5,4,5,5,7,4,5,20,21,22,23,20,21,22,23]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermd %ymm13, %ymm28, %ymm13
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3],xmm6[4],xmm11[4],xmm6[5],xmm11[5],xmm6[6],xmm11[6],xmm6[7],xmm11[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm6 = zmm6[0,1,0,1],mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm26[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm24[2,3,2,3],zmm25[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm22[2,3,2,3],zmm23[2,3,2,3]
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm5 ^ (zmm12 & (zmm7 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vporq %zmm20, %zmm19, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (mem & (zmm15 ^ zmm7))
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT: # zmm7 = zmm7[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm12[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm10 & (zmm0 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm24 ^ (zmm10 & (zmm21 ^ zmm24))
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm5 = zmm5[2,3,2,3],mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm12 & (zmm5 ^ zmm7))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm31 = zmm17 ^ (zmm12 & (zmm31 ^ zmm17))
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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 {{.*#+}} zmm23 = zmm30 ^ (zmm7 & (zmm23 ^ zmm30))
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm10 = mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm21 ^ (zmm7 & (zmm16 ^ zmm21))
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm11 = mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[2,3,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm13, %zmm10
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm13 = mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm13
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm10 ^ (zmm7 & (zmm13 ^ zmm10))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm12, %zmm9 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm7 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm11 ^ (zmm7 & (zmm12 ^ zmm11))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm4))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 | (zmm3 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm7))
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm29 ^ (mem & (zmm4 ^ zmm29))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm23))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm25 ^ (mem & (zmm5 ^ zmm25))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (mem & (zmm5 ^ zmm21))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm13))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm16 ^ (mem & (zmm6 ^ zmm16))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm4 ^ (mem & (zmm30 ^ zmm4))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (mem & (zmm30 ^ zmm16))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm1 ^ (mem & (zmm9 ^ zmm1))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (mem & (zmm9 ^ zmm31))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm1 = zmm1[0,1,2,3],mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 | (zmm1 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm12))
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm8))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm18 ^ (mem & (zmm2 ^ zmm18))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 256(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 192(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 384(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 64(%rax)
-; AVX512DQ-FCP-NEXT: addq $1416, %rsp # imm = 0x588
+; AVX512DQ-FCP-NEXT: addq $1448, %rsp # imm = 0x5A8
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
; AVX512BW-LABEL: store_i8_stride7_vf64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm0
-; AVX512BW-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512BW-NEXT: vmovdqa 32(%rdi), %ymm13
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm13[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm3 = [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: movl $338170920, %r11d # imm = 0x14281428
-; AVX512BW-NEXT: kmovd %r11d, %k2
-; AVX512BW-NEXT: vpshufb %ymm3, %ymm2, %ymm1 {%k2}
-; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
-; AVX512BW-NEXT: vpshufb %ymm8, %ymm2, %ymm4
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm9 = [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 %ymm9, %ymm13, %ymm5
+; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512BW-NEXT: vmovdqa64 32(%rdx), %ymm16
+; 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, %ymm16, %ymm2
+; AVX512BW-NEXT: vmovdqa64 32(%rcx), %ymm17
+; AVX512BW-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]
+; AVX512BW-NEXT: vpshufb %ymm1, %ymm17, %ymm4
+; AVX512BW-NEXT: vpor %ymm2, %ymm4, %ymm2
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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 %ymm4, %ymm16, %ymm4
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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 %ymm5, %ymm17, %ymm5
; AVX512BW-NEXT: vpor %ymm4, %ymm5, %ymm4
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[2,3,2,3],zmm1[2,3,2,3]
-; AVX512BW-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm4 = [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 %ymm4, %ymm14, %ymm5
-; AVX512BW-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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]
-; AVX512BW-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512BW-NEXT: vpshufb %ymm10, %ymm15, %ymm6
-; AVX512BW-NEXT: vpor %ymm5, %ymm6, %ymm6
-; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [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 %ymm5, %ymm15, %ymm7
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT: movl $676341840, %r11d # imm = 0x28502850
-; AVX512BW-NEXT: kmovd %r11d, %k3
-; AVX512BW-NEXT: vmovdqu8 %ymm11, %ymm7 {%k3}
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[2,3,2,3],zmm7[2,3,2,3]
-; AVX512BW-NEXT: movabsq $1742999440035548184, %r11 # imm = 0x183060C183060C18
-; AVX512BW-NEXT: kmovq %r11, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm1 {%k1}
-; AVX512BW-NEXT: vmovdqa64 32(%r9), %ymm16
-; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128]
-; AVX512BW-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512BW-NEXT: vpshufb %ymm6, %ymm16, %ymm11
-; AVX512BW-NEXT: vmovdqa64 32(%r8), %ymm19
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
-; AVX512BW-NEXT: vpshufb %ymm7, %ymm19, %ymm12
-; AVX512BW-NEXT: vpor %ymm11, %ymm12, %ymm11
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [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 %ymm17, %ymm19, %ymm12
-; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm18 = [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]
-; AVX512BW-NEXT: # ymm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpshufb %ymm18, %ymm16, %ymm20
-; AVX512BW-NEXT: vporq %ymm12, %ymm20, %ymm12
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm12[2,3,2,3],zmm11[2,3,2,3]
-; AVX512BW-NEXT: movabsq $6971997760142192736, %r11 # imm = 0x60C183060C183060
-; AVX512BW-NEXT: kmovq %r11, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm1 {%k1}
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm11 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
-; AVX512BW-NEXT: vpermw %zmm0, %zmm11, %zmm11
-; AVX512BW-NEXT: movabsq $-9150747060186627967, %r11 # imm = 0x8102040810204081
-; AVX512BW-NEXT: kmovq %r11, %k4
-; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm1 {%k4}
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm14, %ymm20
-; AVX512BW-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]
-; AVX512BW-NEXT: vpshufb %ymm12, %ymm15, %ymm21
-; AVX512BW-NEXT: vporq %ymm20, %ymm21, %ymm20
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm22 = [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]
-; AVX512BW-NEXT: vpshufb %ymm22, %ymm14, %ymm14
-; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm23 = [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]
-; AVX512BW-NEXT: # ymm23 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpshufb %ymm23, %ymm15, %ymm15
-; AVX512BW-NEXT: vpor %ymm14, %ymm15, %ymm14
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm14, %zmm20, %zmm14
-; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm13[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: 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: vpshufb %ymm25, %ymm2, %ymm15 {%k3}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm13, %ymm13
-; AVX512BW-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-NEXT: vpshufb %ymm21, %ymm2, %ymm2
-; AVX512BW-NEXT: vpor %ymm2, %ymm13, %ymm2
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm15, %zmm2, %zmm2
-; AVX512BW-NEXT: movabsq $3485998880071096368, %r11 # imm = 0x3060C183060C1830
-; AVX512BW-NEXT: kmovq %r11, %k4
-; AVX512BW-NEXT: vmovdqu8 %zmm14, %zmm2 {%k4}
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm16, %ymm15
-; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm14 = [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 %ymm14, %ymm19, %ymm24
-; AVX512BW-NEXT: vporq %ymm15, %ymm24, %ymm15
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [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]
-; AVX512BW-NEXT: vpshufb %ymm26, %ymm19, %ymm19
-; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm28 = [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]
-; AVX512BW-NEXT: # ymm28 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpshufb %ymm28, %ymm16, %ymm16
-; AVX512BW-NEXT: vporq %ymm19, %ymm16, %ymm16
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,3,2,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
-; AVX512BW-NEXT: vmovdqa64 32(%r10), %ymm16
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm29 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
-; AVX512BW-NEXT: vpermw %ymm16, %ymm29, %ymm19
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm16 = ymm16[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: vinserti64x4 $1, %ymm19, %zmm16, %zmm16
-; AVX512BW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
-; AVX512BW-NEXT: kmovq %r10, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm16, %zmm15 {%k5}
-; AVX512BW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512BW-NEXT: kmovq %r10, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm15, %zmm2 {%k5}
-; AVX512BW-NEXT: vmovdqa (%rdx), %ymm15
-; AVX512BW-NEXT: vpshufb %ymm4, %ymm15, %ymm4
-; AVX512BW-NEXT: vmovdqa64 (%rcx), %ymm16
-; AVX512BW-NEXT: vpshufb %ymm10, %ymm16, %ymm10
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512BW-NEXT: vmovdqa64 32(%rsi), %ymm24
+; AVX512BW-NEXT: vmovdqa64 32(%rdi), %ymm23
+; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm23[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,1,1,4,4,5,5]
+; AVX512BW-NEXT: movl $676341840, %r10d # imm = 0x28502850
+; AVX512BW-NEXT: kmovd %r10d, %k1
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm4 {%k1} = ymm24[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm7 = [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 %ymm7, %ymm23, %ymm5
+; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm8 = [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 %ymm8, %ymm24, %ymm6
+; AVX512BW-NEXT: vpor %ymm5, %ymm6, %ymm5
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm5
+; AVX512BW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
+; AVX512BW-NEXT: kmovq %r10, %k2
+; AVX512BW-NEXT: vmovdqu8 %zmm2, %zmm5 {%k2}
+; AVX512BW-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm2 = [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 %ymm2, %ymm4, %ymm9
+; AVX512BW-NEXT: vmovdqa 32(%r8), %ymm10
+; AVX512BW-NEXT: vmovdqa {{.*#+}} ymm6 = [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 %ymm6, %ymm10, %ymm11
+; AVX512BW-NEXT: vpor %ymm9, %ymm11, %ymm9
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [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 %ymm11, %ymm10, %ymm10
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [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 %ymm11, %ymm4, %ymm4
; AVX512BW-NEXT: vpor %ymm4, %ymm10, %ymm4
-; AVX512BW-NEXT: vpshufb %ymm22, %ymm15, %ymm10
-; AVX512BW-NEXT: vpshufb %ymm23, %ymm16, %ymm19
-; AVX512BW-NEXT: vporq %ymm10, %ymm19, %ymm10
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm10[2,3,2,3],zmm4[2,3,2,3]
-; AVX512BW-NEXT: vmovdqa64 (%rsi), %ymm24
-; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm27
-; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm27[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,1,1,4,4,5,5]
-; AVX512BW-NEXT: vpshufb %ymm25, %ymm24, %ymm10 {%k3}
-; AVX512BW-NEXT: vpshufb %ymm8, %ymm24, %ymm8
-; AVX512BW-NEXT: vpshufb %ymm9, %ymm27, %ymm9
-; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm10[2,3,2,3],zmm8[2,3,2,3]
-; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm4 {%k1}
-; AVX512BW-NEXT: vmovdqa64 (%r8), %ymm19
-; AVX512BW-NEXT: vpshufb %ymm17, %ymm19, %ymm8
-; AVX512BW-NEXT: vmovdqa64 (%r9), %ymm22
-; AVX512BW-NEXT: vpshufb %ymm18, %ymm22, %ymm9
-; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512BW-NEXT: vpshufb %ymm26, %ymm19, %ymm9
-; AVX512BW-NEXT: vpshufb %ymm28, %ymm22, %ymm10
-; AVX512BW-NEXT: vpor %ymm9, %ymm10, %ymm9
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
-; AVX512BW-NEXT: vpermw %zmm0, %zmm29, %zmm9
-; AVX512BW-NEXT: movabsq $1161999626690365456, %r10 # imm = 0x1020408102040810
-; AVX512BW-NEXT: kmovq %r10, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm8 {%k5}
-; AVX512BW-NEXT: movabsq $2033499346708139548, %r10 # imm = 0x1C3870E1C3870E1C
-; AVX512BW-NEXT: kmovq %r10, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm8, %zmm4 {%k5}
-; AVX512BW-NEXT: vpshufb %ymm11, %ymm15, %ymm8
-; AVX512BW-NEXT: vpshufb %ymm12, %ymm16, %ymm9
-; AVX512BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512BW-NEXT: vmovdqa64 32(%rdx), %xmm17
-; AVX512BW-NEXT: vmovdqa (%rcx), %xmm10
-; AVX512BW-NEXT: vmovdqa64 32(%rcx), %xmm18
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm23 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm11, %xmm11
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm26
-; AVX512BW-NEXT: vpshufb %ymm20, %ymm27, %ymm8
-; AVX512BW-NEXT: vpshufb %ymm21, %ymm24, %ymm11
-; AVX512BW-NEXT: vpor %ymm8, %ymm11, %ymm8
-; AVX512BW-NEXT: vmovdqa (%rdi), %xmm11
-; AVX512BW-NEXT: vmovdqa64 32(%rdi), %xmm20
-; AVX512BW-NEXT: vmovdqa (%rsi), %xmm12
-; AVX512BW-NEXT: vmovdqa64 32(%rsi), %xmm21
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm28 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-NEXT: vpshufb %xmm25, %xmm28, %xmm28
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm28, %zmm8
-; AVX512BW-NEXT: movabsq $435749860008887046, %rcx # imm = 0x60C183060C18306
-; AVX512BW-NEXT: kmovq %rcx, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm26, %zmm8 {%k5}
-; AVX512BW-NEXT: vpshufb %ymm13, %ymm22, %ymm13
-; AVX512BW-NEXT: vpshufb %ymm14, %ymm19, %ymm14
-; AVX512BW-NEXT: vporq %ymm13, %ymm14, %ymm30
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm13
-; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm26
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm14
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm9, %zmm4
+; AVX512BW-NEXT: vmovdqa 32(%rax), %ymm9
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-NEXT: vpermw %ymm9, %ymm10, %ymm10
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[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: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512BW-NEXT: movabsq $145249953336295682, %rax # imm = 0x204081020408102
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm4 {%k3}
+; AVX512BW-NEXT: movabsq $-4357498600088870461, %rax # imm = 0xC3870E1C3870E1C3
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm4, %zmm5 {%k3}
+; AVX512BW-NEXT: vmovdqa (%rdi), %ymm14
+; AVX512BW-NEXT: vpshufb %ymm7, %ymm14, %ymm4
+; AVX512BW-NEXT: vmovdqa (%rsi), %ymm15
+; AVX512BW-NEXT: vpshufb %ymm8, %ymm15, %ymm7
+; AVX512BW-NEXT: vpor %ymm4, %ymm7, %ymm4
+; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm25
+; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512BW-NEXT: vmovdqa64 (%rsi), %xmm26
+; AVX512BW-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT: vpshufb %ymm0, %ymm18, %ymm0
+; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT: vpshufb %ymm1, %ymm19, %ymm1
+; AVX512BW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm18[8],xmm19[9],xmm18[9],xmm19[10],xmm18[10],xmm19[11],xmm18[11],xmm19[12],xmm18[12],xmm19[13],xmm18[13],xmm19[14],xmm18[14],xmm19[15],xmm18[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512BW-NEXT: movabsq $435749860008887046, %rax # imm = 0x60C183060C18306
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k3}
+; AVX512BW-NEXT: vmovdqa64 (%r9), %ymm20
+; AVX512BW-NEXT: vpshufb %ymm2, %ymm20, %ymm0
+; AVX512BW-NEXT: vmovdqa64 (%r8), %ymm21
+; AVX512BW-NEXT: vpshufb %ymm6, %ymm21, %ymm1
+; AVX512BW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512BW-NEXT: vmovdqa64 (%r9), %xmm30
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm30[8],xmm2[9],xmm30[9],xmm2[10],xmm30[10],xmm2[11],xmm30[11],xmm2[12],xmm30[12],xmm2[13],xmm30[13],xmm2[14],xmm30[14],xmm2[15],xmm30[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
+; AVX512BW-NEXT: vpermw %zmm3, %zmm1, %zmm1
+; AVX512BW-NEXT: movabsq $2323999253380730912, %rax # imm = 0x2040810204081020
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
+; AVX512BW-NEXT: movabsq $4066998693416279096, %rax # imm = 0x3870E1C3870E1C38
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k3}
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm14[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT: vpbroadcastd {{.*#+}} ymm7 = [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: movl $338170920, %eax # imm = 0x14281428
+; AVX512BW-NEXT: kmovd %eax, %k3
+; AVX512BW-NEXT: vpshufb %ymm7, %ymm15, %ymm0 {%k3}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm6[0,1,0,1]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm18[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,2,3,3,4,6,7,7]
+; AVX512BW-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]
+; AVX512BW-NEXT: vpshufb %ymm1, %ymm19, %ymm9
+; AVX512BW-NEXT: vmovdqu8 %ymm6, %ymm9 {%k1}
+; AVX512BW-NEXT: vmovdqa64 32(%rcx), %xmm29
+; AVX512BW-NEXT: vmovdqa64 32(%rdx), %xmm31
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm31[0],xmm29[0],xmm31[1],xmm29[1],xmm31[2],xmm29[2],xmm31[3],xmm29[3],xmm31[4],xmm29[4],xmm31[5],xmm29[5],xmm31[6],xmm29[6],xmm31[7],xmm29[7]
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm6, %xmm6
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm9[2,3,2,3],zmm6[0,1,0,1]
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm22 {%k2}
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm20[27],zero,zero,zero,zero,ymm20[30],zero,ymm20[28],zero,zero,zero,zero,ymm20[31],zero
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm21[27],zero,zero,zero,zero,ymm21[30],zero,ymm21[28],zero,zero,zero,zero,ymm21[31],zero,ymm21[29]
+; AVX512BW-NEXT: vpor %ymm0, %ymm6, %ymm0
+; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm27
; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm28
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm31 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm29 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm31, %xmm31
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,1,0,1]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm30, %zmm31, %zmm30
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm31 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
-; AVX512BW-NEXT: vpermw %zmm0, %zmm31, %zmm31
-; AVX512BW-NEXT: movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
-; AVX512BW-NEXT: kmovq %rcx, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm31, %zmm30 {%k5}
-; AVX512BW-NEXT: movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
-; AVX512BW-NEXT: kmovq %rcx, %k5
-; AVX512BW-NEXT: vmovdqu8 %zmm30, %zmm8 {%k5}
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm27 = ymm27[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm27 = ymm27[2,2,3,3,6,6,7,7]
-; AVX512BW-NEXT: vpshufb %ymm3, %ymm24, %ymm27 {%k2}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm20[0],xmm21[0],xmm20[1],xmm21[1],xmm20[2],xmm21[2],xmm20[3],xmm21[3],xmm20[4],xmm21[4],xmm20[5],xmm21[5],xmm20[6],xmm21[6],xmm20[7],xmm21[7]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm24 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm3, %xmm3
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm27[2,3,2,3],zmm3[0,1,0,1]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3],xmm17[4],xmm18[4],xmm17[5],xmm18[5],xmm17[6],xmm18[6],xmm17[7],xmm18[7]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm27 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-NEXT: vpshufb %xmm27, %xmm3, %xmm3
-; AVX512BW-NEXT: vpshufb %ymm5, %ymm16, %ymm5
-; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,2,3,3,4,6,7,7]
-; AVX512BW-NEXT: vmovdqu8 %ymm15, %ymm5 {%k3}
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm5[2,3,2,3],zmm3[0,1,0,1]
-; AVX512BW-NEXT: vmovdqu8 %zmm30, %zmm3 {%k4}
-; AVX512BW-NEXT: vpshufb %ymm6, %ymm22, %ymm5
-; AVX512BW-NEXT: vpshufb %ymm7, %ymm19, %ymm6
-; AVX512BW-NEXT: vpor %ymm5, %ymm6, %ymm6
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm28[0],xmm26[0],xmm28[1],xmm26[1],xmm28[2],xmm26[2],xmm28[3],xmm26[3],xmm28[4],xmm26[4],xmm28[5],xmm26[5],xmm28[6],xmm26[6],xmm28[7],xmm26[7]
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm5 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm7, %xmm7
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[2,3,2,3],zmm7[0,1,0,1]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [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 %zmm0, %zmm7, %zmm7
-; AVX512BW-NEXT: movabsq $580999813345182728, %rcx # imm = 0x810204081020408
-; AVX512BW-NEXT: kmovq %rcx, %k2
-; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm6 {%k2}
-; AVX512BW-NEXT: movabsq $1016749673354069774, %rcx # imm = 0xE1C3870E1C3870E
-; AVX512BW-NEXT: kmovq %rcx, %k2
-; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k2}
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm18, %xmm7
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm15 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm17, %xmm16
-; AVX512BW-NEXT: vporq %xmm7, %xmm16, %xmm7
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm16, %xmm16
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm7[0,1,0,1],zmm16[0,1,0,1]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm16 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512BW-NEXT: vpshufb %xmm16, %xmm21, %xmm17
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm18 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512BW-NEXT: vpshufb %xmm18, %xmm20, %xmm19
-; AVX512BW-NEXT: vporq %xmm17, %xmm19, %xmm17
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
-; AVX512BW-NEXT: vpshufb %xmm25, %xmm19, %xmm19
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm17[0,1,0,1],zmm19[0,1,0,1]
-; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm17 {%k1}
-; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm7 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm26, %xmm19
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm20 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm28, %xmm21
-; AVX512BW-NEXT: vporq %xmm19, %xmm21, %xmm19
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm28[8],xmm26[8],xmm28[9],xmm26[9],xmm28[10],xmm26[10],xmm28[11],xmm26[11],xmm28[12],xmm26[12],xmm28[13],xmm26[13],xmm28[14],xmm26[14],xmm28[15],xmm26[15]
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm21, %xmm21
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,0,1],zmm21[0,1,0,1]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm21 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
-; AVX512BW-NEXT: vpermw %zmm0, %zmm21, %zmm21
-; AVX512BW-NEXT: movabsq $290499906672591364, %rcx # imm = 0x408102040810204
-; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm21, %zmm19 {%k1}
-; AVX512BW-NEXT: movabsq $-8714997200177740921, %rcx # imm = 0x870E1C3870E1C387
-; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm19, %zmm17 {%k1}
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm10, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm9, %xmm15
-; AVX512BW-NEXT: vpor %xmm6, %xmm15, %xmm6
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512BW-NEXT: vpshufb %xmm27, %xmm9, %xmm9
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,0,1],zmm6[0,1,0,1]
-; AVX512BW-NEXT: vpshufb %xmm16, %xmm12, %xmm9
-; AVX512BW-NEXT: vpshufb %xmm18, %xmm11, %xmm10
-; AVX512BW-NEXT: vpor %xmm9, %xmm10, %xmm9
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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-NEXT: vpshufb %xmm24, %xmm10, %xmm10
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[0,1,0,1],zmm9[0,1,0,1]
-; AVX512BW-NEXT: movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
-; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm9 {%k1}
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm13, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm14, %xmm7
-; AVX512BW-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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: vpshufb %xmm5, %xmm7, %xmm5
-; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,0,1],zmm6[0,1,0,1]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
-; AVX512BW-NEXT: vpermw %zmm0, %zmm6, %zmm0
-; AVX512BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
-; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm5 {%k1}
-; AVX512BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
-; AVX512BW-NEXT: kmovq %rcx, %k1
-; AVX512BW-NEXT: vmovdqu8 %zmm5, %zmm9 {%k1}
-; AVX512BW-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm17, 256(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm28[0],xmm27[0],xmm28[1],xmm27[1],xmm28[2],xmm27[2],xmm28[3],xmm27[3],xmm28[4],xmm27[4],xmm28[5],xmm27[5],xmm28[6],xmm27[6],xmm28[7],xmm27[7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm6[0,1,0,1]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [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 %zmm3, %zmm6, %zmm6
+; AVX512BW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512BW-NEXT: kmovq %rax, %k2
+; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm0 {%k2}
+; AVX512BW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512BW-NEXT: kmovq %rax, %k2
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm22 {%k2}
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm23[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
+; AVX512BW-NEXT: vpshufb %ymm7, %ymm24, %ymm0 {%k3}
+; AVX512BW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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: # ymm7 = mem[0,1,0,1]
+; AVX512BW-NEXT: vpshufb %ymm7, %ymm24, %ymm6
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [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 %ymm24, %ymm23, %ymm23
+; AVX512BW-NEXT: vporq %ymm6, %ymm23, %ymm6
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm6[2,3,2,3],zmm0[2,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[24,25],zero,ymm16[23],zero,ymm16[21,22,23,26],zero,ymm16[24],zero,ymm16[28,29,26,27]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm17[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm17[25],zero,ymm17[23],zero,zero,zero,zero,ymm17[26],zero,ymm17[24],zero,zero,zero,zero
+; AVX512BW-NEXT: vpor %ymm0, %ymm6, %ymm0
+; AVX512BW-NEXT: vpshufb %ymm1, %ymm17, %ymm1
+; AVX512BW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,2,3,3,4,6,7,7]
+; AVX512BW-NEXT: vmovdqu8 %ymm6, %ymm1 {%k1}
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm1[2,3,2,3]
+; AVX512BW-NEXT: movabsq $1742999440035548184, %rax # imm = 0x183060C183060C18
+; AVX512BW-NEXT: kmovq %rax, %k2
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm23 {%k2}
+; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm0
+; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm1
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[4,5,6,7],zmm1[4,5,6,7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm6 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm6[23],zero,zmm6[23,24,25,26],zero,zmm6[24],zero,zmm6[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm6[59],zero,zero,zero,zero,zmm6[62],zero,zmm6[60],zero,zero,zero,zero,zmm6[63],zero
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[4,5,6,7],zmm0[4,5,6,7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm16[25],zero,zmm16[23],zero,zero,zero,zero,zmm16[26],zero,zmm16[24],zero,zero,zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[59],zero,zero,zero,zero,zmm16[62],zero,zmm16[60],zero,zero,zero,zero,zmm16[63],zero,zmm16[61]
+; AVX512BW-NEXT: vporq %zmm6, %zmm16, %zmm6
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512BW-NEXT: kmovq %rax, %k2
+; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm23 {%k2}
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
+; AVX512BW-NEXT: vpermw %zmm3, %zmm6, %zmm6
+; AVX512BW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm6, %zmm23 {%k3}
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm17 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm26, %xmm16
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm25, %xmm13
+; AVX512BW-NEXT: vporq %xmm16, %xmm13, %xmm13
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm25[0],xmm26[0],xmm25[1],xmm26[1],xmm25[2],xmm26[2],xmm25[3],xmm26[3],xmm25[4],xmm26[4],xmm25[5],xmm26[5],xmm25[6],xmm26[6],xmm25[7],xmm26[7]
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm16, %xmm12
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm13[0,1,0,1]
+; AVX512BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm19, %xmm12
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm13 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-NEXT: vpshufb %xmm13, %xmm18, %xmm26
+; AVX512BW-NEXT: vporq %xmm12, %xmm26, %xmm12
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm18[0],xmm19[0],xmm18[1],xmm19[1],xmm18[2],xmm19[2],xmm18[3],xmm19[3],xmm18[4],xmm19[4],xmm18[5],xmm19[5],xmm18[6],xmm19[6],xmm18[7],xmm19[7]
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm26, %xmm11
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm11[0,1,0,1],zmm12[0,1,0,1]
+; AVX512BW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm11, %zmm16 {%k3}
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm11 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm30, %xmm26
+; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm12 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm2, %xmm9
+; AVX512BW-NEXT: vporq %xmm26, %xmm9, %xmm9
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3],xmm2[4],xmm30[4],xmm2[5],xmm30[5],xmm2[6],xmm30[6],xmm2[7],xmm30[7]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm9[0,1,0,1]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-NEXT: vpermw %zmm3, %zmm9, %zmm9
+; AVX512BW-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm2 {%k3}
+; AVX512BW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
+; AVX512BW-NEXT: kmovq %rax, %k3
+; AVX512BW-NEXT: vmovdqu8 %zmm2, %zmm16 {%k3}
+; AVX512BW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm14[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,1,1,4,4,5,5]
+; AVX512BW-NEXT: vpshufb {{.*#+}} ymm2 {%k1} = ymm15[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512BW-NEXT: vpshufb %ymm7, %ymm15, %ymm7
+; AVX512BW-NEXT: vpshufb %ymm24, %ymm14, %ymm9
+; AVX512BW-NEXT: vpor %ymm7, %ymm9, %ymm7
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[2,3,2,3],zmm7[2,3,2,3]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm18[18],zero,zmm18[18,19,20,21],zero,zmm18[19],zero,zmm18[25,26,27,22],zero,zmm18[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm18[56,57],zero,zmm18[55],zero,zmm18[53,54,55,58],zero,zmm18[56],zero,zmm18[60,61,58,59]
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm19[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm19[18],zero,zero,zero,zero,zmm19[21],zero,zmm19[19],zero,zero,zero,zero,zmm19[22],zero,zmm19[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm19[57],zero,zmm19[55],zero,zero,zero,zero,zmm19[58],zero,zmm19[56],zero,zero,zero,zero
+; AVX512BW-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: vmovdqu8 %zmm2, %zmm7 {%k2}
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm0
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[20],zero,zmm0[18],zero,zmm0[20,21,20,21],zero,zmm0[19],zero,zmm0[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[56,57,56,57],zero,zmm0[55],zero,zmm0[55,56,57,58],zero,zmm0[56],zero,zmm0[62,63]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm1, %zmm1
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm1[20],zero,zmm1[18],zero,zero,zero,zero,zmm1[21],zero,zmm1[19],zero,zero,zero,zero,zmm1[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm1[57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero
+; AVX512BW-NEXT: vporq %zmm0, %zmm1, %zmm0
+; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-NEXT: vpermw %zmm3, %zmm1, %zmm1
+; AVX512BW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
+; AVX512BW-NEXT: kmovq %rax, %k1
+; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512BW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512BW-NEXT: kmovq %rax, %k1
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm7 {%k1}
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm29, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm13, %xmm31, %xmm1
+; AVX512BW-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm29[8],xmm31[8],xmm29[9],xmm31[9],xmm29[10],xmm31[10],xmm29[11],xmm31[11],xmm29[12],xmm31[12],xmm29[13],xmm31[13],xmm29[14],xmm31[14],xmm29[15],xmm31[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,0,1],zmm1[0,1,0,1]
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm10, %xmm1
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm8, %xmm2
+; AVX512BW-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,0,1],zmm2[0,1,0,1]
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
+; AVX512BW-NEXT: vpshufb %xmm11, %xmm27, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm28, %xmm2
+; AVX512BW-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm28[8],xmm27[8],xmm28[9],xmm27[9],xmm28[10],xmm27[10],xmm28[11],xmm27[11],xmm28[12],xmm27[12],xmm28[13],xmm27[13],xmm28[14],xmm27[14],xmm28[15],xmm27[15]
+; AVX512BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,0,1],zmm2[0,1,0,1]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm2 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
+; AVX512BW-NEXT: vpermw %zmm3, %zmm2, %zmm2
+; AVX512BW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
+; AVX512BW-NEXT: kmovq %rax, %k1
+; AVX512BW-NEXT: vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512BW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
+; AVX512BW-NEXT: kmovq %rax, %k1
+; AVX512BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-NEXT: vmovdqa64 %zmm1, 256(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm16, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm23, 384(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm22, 192(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm4, 64(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
; AVX512BW-FCP-LABEL: store_i8_stride7_vf64:
; AVX512BW-FCP: # %bb.0:
+; AVX512BW-FCP-NEXT: subq $200, %rsp
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm0
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128]
-; AVX512BW-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128]
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm5
-; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm5, %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [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 %ymm7, %ymm13, %ymm5
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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]
-; AVX512BW-FCP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm6
-; AVX512BW-FCP-NEXT: vpor %ymm5, %ymm6, %ymm5
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[2,3,2,3],zmm1[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
-; AVX512BW-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm1
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm16
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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]
-; AVX512BW-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm16, %ymm6
-; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm6, %ymm1
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm18 = [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-FCP-NEXT: # ymm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm18, %ymm16, %ymm6
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [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 %ymm19, %ymm15, %ymm11
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm8
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm7
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm4
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm17
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm3
+; AVX512BW-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]
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
+; AVX512BW-FCP-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]
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm11
; AVX512BW-FCP-NEXT: vpor %ymm6, %ymm11, %ymm6
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm6[2,3,2,3],zmm1[2,3,2,3]
-; AVX512BW-FCP-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
-; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %ymm17
-; AVX512BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128]
-; AVX512BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm17, %ymm11
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm22
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
-; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm22, %ymm12
-; AVX512BW-FCP-NEXT: vpor %ymm11, %ymm12, %ymm11
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm22, %ymm12
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm21 = [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]
-; AVX512BW-FCP-NEXT: # ymm21 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm21, %ymm17, %ymm14
-; AVX512BW-FCP-NEXT: vpor %ymm12, %ymm14, %ymm12
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm12[2,3,2,3],zmm11[2,3,2,3]
-; AVX512BW-FCP-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
-; AVX512BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm1 {%k1}
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm11 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
-; AVX512BW-FCP-NEXT: vpermw %zmm0, %zmm11, %zmm11
-; AVX512BW-FCP-NEXT: movabsq $-9150747060186627967, %r10 # imm = 0x8102040810204081
-; AVX512BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm1 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm13, %ymm14
-; AVX512BW-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]
-; AVX512BW-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm23
-; AVX512BW-FCP-NEXT: vporq %ymm14, %ymm23, %ymm14
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm24 = [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]
-; AVX512BW-FCP-NEXT: vpshufb %ymm24, %ymm13, %ymm13
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm25 = [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]
-; AVX512BW-FCP-NEXT: # ymm25 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm25, %ymm2, %ymm2
-; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm13, %ymm2
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm23
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm15, %ymm2
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [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 %ymm14, %ymm16, %ymm26
-; AVX512BW-FCP-NEXT: vporq %ymm2, %ymm26, %ymm2
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm26 = [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]
-; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm15, %ymm15
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm27 = [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]
-; AVX512BW-FCP-NEXT: # ymm27 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm27, %ymm16, %ymm16
-; AVX512BW-FCP-NEXT: vporq %ymm15, %ymm16, %ymm15
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm2, %zmm2
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm3, %ymm3
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm10, %ymm10
+; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm10, %ymm3
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm11
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm6
+; AVX512BW-FCP-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]
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm10
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm12
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm12, %ymm14
+; AVX512BW-FCP-NEXT: vpor %ymm10, %ymm14, %ymm14
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm6, %ymm6
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm12, %ymm12
+; AVX512BW-FCP-NEXT: vpor %ymm6, %ymm12, %ymm6
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm14, %zmm10
; AVX512BW-FCP-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm2 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [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 %ymm15, %ymm17, %ymm23
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm16 = [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 %ymm16, %ymm22, %ymm28
-; AVX512BW-FCP-NEXT: vporq %ymm23, %ymm28, %ymm23
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm28 = [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]
-; AVX512BW-FCP-NEXT: vpshufb %ymm28, %ymm22, %ymm22
-; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm29 = [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]
-; AVX512BW-FCP-NEXT: # ymm29 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-FCP-NEXT: vpshufb %ymm29, %ymm17, %ymm17
-; AVX512BW-FCP-NEXT: vporq %ymm22, %ymm17, %ymm17
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm23, %zmm17
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rax), %ymm22
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm30 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
-; AVX512BW-FCP-NEXT: vpermw %ymm22, %ymm30, %ymm23
-; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm22 = ymm22[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: vinserti64x4 $1, %ymm23, %zmm22, %zmm22
+; AVX512BW-FCP-NEXT: kmovq %r10, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm10 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm14
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [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 %ymm19, %ymm14, %ymm11
+; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %ymm12
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm12, %ymm15
+; AVX512BW-FCP-NEXT: vpor %ymm11, %ymm15, %ymm15
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [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 %ymm6, %ymm12, %ymm16
+; 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, %ymm14, %ymm14
+; AVX512BW-FCP-NEXT: vporq %ymm16, %ymm14, %ymm14
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm15, %zmm14
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm16 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512BW-FCP-NEXT: vpermw %ymm15, %ymm16, %ymm16
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[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: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
; AVX512BW-FCP-NEXT: movabsq $145249953336295682, %rax # imm = 0x204081020408102
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm17 {%k3}
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm14 {%k2}
; AVX512BW-FCP-NEXT: movabsq $-4357498600088870461, %rax # imm = 0xC3870E1C3870E1C3
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm2 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm17
-; AVX512BW-FCP-NEXT: vpshufb %ymm18, %ymm17, %ymm18
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
-; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm22, %ymm19
-; AVX512BW-FCP-NEXT: vporq %ymm18, %ymm19, %ymm18
-; AVX512BW-FCP-NEXT: vpshufb %ymm26, %ymm22, %ymm19
-; AVX512BW-FCP-NEXT: vpshufb %ymm27, %ymm17, %ymm23
-; AVX512BW-FCP-NEXT: vporq %ymm19, %ymm23, %ymm19
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[2,3,2,3],zmm18[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm18
-; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm18, %ymm7
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm23
-; AVX512BW-FCP-NEXT: vpshufb %ymm10, %ymm23, %ymm10
-; AVX512BW-FCP-NEXT: vpor %ymm7, %ymm10, %ymm7
-; AVX512BW-FCP-NEXT: vpshufb %ymm24, %ymm18, %ymm10
-; AVX512BW-FCP-NEXT: vpshufb %ymm25, %ymm23, %ymm24
-; AVX512BW-FCP-NEXT: vporq %ymm10, %ymm24, %ymm10
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm10[2,3,2,3],zmm7[2,3,2,3]
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm7 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %ymm19
-; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm19, %ymm10
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %ymm20
-; AVX512BW-FCP-NEXT: vpshufb %ymm21, %ymm20, %ymm21
-; AVX512BW-FCP-NEXT: vporq %ymm10, %ymm21, %ymm10
-; AVX512BW-FCP-NEXT: vpshufb %ymm28, %ymm19, %ymm21
-; AVX512BW-FCP-NEXT: vpshufb %ymm29, %ymm20, %ymm24
-; AVX512BW-FCP-NEXT: vporq %ymm21, %ymm24, %ymm21
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm21[2,3,2,3],zmm10[2,3,2,3]
-; AVX512BW-FCP-NEXT: vpermw %zmm0, %zmm30, %zmm21
-; AVX512BW-FCP-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm21, %zmm10 {%k3}
-; AVX512BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm7 {%k3}
-; AVX512BW-FCP-NEXT: vpshufb %ymm11, %ymm18, %ymm10
-; AVX512BW-FCP-NEXT: vpshufb %ymm12, %ymm23, %ymm11
-; AVX512BW-FCP-NEXT: vpor %ymm10, %ymm11, %ymm10
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm24, %xmm24
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm24, %zmm25
-; AVX512BW-FCP-NEXT: vpshufb %ymm13, %ymm22, %ymm10
-; AVX512BW-FCP-NEXT: vpshufb %ymm14, %ymm17, %ymm13
-; AVX512BW-FCP-NEXT: vpor %ymm10, %ymm13, %ymm10
-; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm13
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm26 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm24, %xmm26, %xmm26
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm26, %zmm10
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm10 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512BW-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512BW-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm2
+; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm21
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm22
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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: vpshufb {{.*#+}} xmm2 = xmm2[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm16
+; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm16, %ymm2
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm11
+; AVX512BW-FCP-NEXT: vpshufb %ymm13, %ymm11, %ymm3
+; AVX512BW-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm23
+; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm24
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm24[8],xmm23[8],xmm24[9],xmm23[9],xmm24[10],xmm23[10],xmm24[11],xmm23[11],xmm24[12],xmm23[12],xmm24[13],xmm23[13],xmm24[14],xmm23[14],xmm24[15],xmm23[15]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm13
; AVX512BW-FCP-NEXT: movabsq $435749860008887046, %rax # imm = 0x60C183060C18306
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm25, %zmm10 {%k3}
-; AVX512BW-FCP-NEXT: vpshufb %ymm15, %ymm20, %ymm15
-; AVX512BW-FCP-NEXT: vpshufb %ymm16, %ymm19, %ymm16
-; AVX512BW-FCP-NEXT: vporq %ymm15, %ymm16, %ymm26
-; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm15
-; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %xmm16
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm27 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-FCP-NEXT: vpshufb %xmm25, %xmm27, %xmm27
-; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm27 = ymm27[0,1,0,1]
-; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm27, %zmm26
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm27 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
-; AVX512BW-FCP-NEXT: vpermw %zmm0, %zmm27, %zmm27
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm13 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %ymm18
+; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm18, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %ymm19
+; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm19, %ymm2
+; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %xmm25
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %xmm27
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm27[8],xmm25[8],xmm27[9],xmm25[9],xmm27[10],xmm25[10],xmm27[11],xmm25[11],xmm27[12],xmm25[12],xmm27[13],xmm25[13],xmm27[14],xmm25[14],xmm27[15],xmm25[15]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
+; AVX512BW-FCP-NEXT: vpermw %zmm17, %zmm2, %zmm2
; AVX512BW-FCP-NEXT: movabsq $2323999253380730912, %rax # imm = 0x2040810204081020
-; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm27, %zmm26 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm27
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k2}
; AVX512BW-FCP-NEXT: movabsq $4066998693416279096, %rax # imm = 0x3870E1C3870E1C38
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm13 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[4,5,6,7],zmm4[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: vshufi64x2 {{.*#+}} zmm2 = zmm4[4,5,6,7],zmm5[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: vporq %zmm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm7[4,5,6,7],zmm8[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: vshufi64x2 {{.*#+}} zmm3 = zmm8[4,5,6,7],zmm7[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: vporq %zmm2, %zmm3, %zmm2
+; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm26
+; 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, %rax # imm = 0x183060C183060C18
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm0[4,5,6,7],zmm26[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: vshufi64x2 {{.*#+}} zmm2 = zmm26[4,5,6,7],zmm0[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: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
+; AVX512BW-FCP-NEXT: vpermw %zmm17, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm26, %zmm10 {%k3}
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm26
-; AVX512BW-FCP-NEXT: vpshufb %ymm8, %ymm22, %ymm8
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm22
-; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm17, %ymm9
-; AVX512BW-FCP-NEXT: vpor %ymm8, %ymm9, %ymm9
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm26[0],xmm22[0],xmm26[1],xmm22[1],xmm26[2],xmm22[2],xmm26[3],xmm22[3],xmm26[4],xmm22[4],xmm26[5],xmm22[5],xmm26[6],xmm22[6],xmm26[7],xmm22[7]
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm17, %xmm17
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm9[2,3,2,3],zmm17[0,1,0,1]
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm28
-; AVX512BW-FCP-NEXT: vpshufb %ymm3, %ymm23, %ymm3
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm23
-; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm18, %ymm4
-; AVX512BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm27[0],xmm28[0],xmm27[1],xmm28[1],xmm27[2],xmm28[2],xmm27[3],xmm28[3],xmm27[4],xmm28[4],xmm27[5],xmm28[5],xmm27[6],xmm28[6],xmm27[7],xmm28[7]
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-FCP-NEXT: vpshufb %xmm9, %xmm4, %xmm4
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[2,3,2,3],zmm4[0,1,0,1]
-; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm18
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm3 {%k2}
-; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm20, %ymm4
-; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm19, %ymm5
-; AVX512BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm5
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm18[0],xmm23[0],xmm18[1],xmm23[1],xmm18[2],xmm23[2],xmm18[3],xmm23[3],xmm18[4],xmm23[4],xmm18[5],xmm23[5],xmm18[6],xmm23[6],xmm18[7],xmm23[7]
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm4, %xmm6, %xmm6
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[2,3,2,3],zmm6[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [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 %zmm0, %zmm6, %zmm6
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[28,29,30],zero,ymm16[28],zero,ymm16[30,31,30,31],zero,ymm16[29],zero,ymm16[31,28,29]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero,zero,zero
+; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm7
+; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm6
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm31 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512BW-FCP-NEXT: vpshufb %xmm31, %xmm3, %xmm3
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm1[2,3,2,3],zmm3[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm29 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29],zero,zero
+; AVX512BW-FCP-NEXT: vporq %ymm1, %ymm29, %ymm29
+; 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: vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512BW-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm29[2,3,2,3],zmm0[0,1,0,1]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm29 {%k1}
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm18[27],zero,zero,zero,zero,ymm18[30],zero,ymm18[28],zero,zero,zero,zero,ymm18[31],zero
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm19[27],zero,zero,zero,zero,ymm19[30],zero,ymm19[28],zero,zero,zero,zero,ymm19[31],zero,ymm19[29]
+; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm3, %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 {{.*#+}} xmm30 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm30, %xmm0, %xmm0
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[2,3,2,3],zmm0[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 %zmm17, %zmm3, %zmm3
; AVX512BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm5 {%k2}
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm3 {%k2}
-; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm28, %xmm5
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm17 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm27, %xmm19
-; AVX512BW-FCP-NEXT: vporq %xmm5, %xmm19, %xmm5
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm28[8],xmm27[8],xmm28[9],xmm27[9],xmm28[10],xmm27[10],xmm28[11],xmm27[11],xmm28[12],xmm27[12],xmm28[13],xmm27[13],xmm28[14],xmm27[14],xmm28[15],xmm27[15]
-; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm19, %xmm19
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm5[0,1,0,1],zmm19[0,1,0,1]
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm20 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm22, %xmm5
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm26, %xmm27
-; AVX512BW-FCP-NEXT: vporq %xmm5, %xmm27, %xmm5
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm22[8],xmm26[8],xmm22[9],xmm26[9],xmm22[10],xmm26[10],xmm22[11],xmm26[11],xmm22[12],xmm26[12],xmm22[13],xmm26[13],xmm22[14],xmm26[14],xmm22[15],xmm26[15]
-; AVX512BW-FCP-NEXT: vpshufb %xmm24, %xmm22, %xmm22
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,0,1],zmm22[0,1,0,1]
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm5 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm23, %xmm22
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512BW-FCP-NEXT: vpshufb %xmm24, %xmm18, %xmm26
-; AVX512BW-FCP-NEXT: vporq %xmm22, %xmm26, %xmm22
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm18[8],xmm23[8],xmm18[9],xmm23[9],xmm18[10],xmm23[10],xmm18[11],xmm23[11],xmm18[12],xmm23[12],xmm18[13],xmm23[13],xmm18[14],xmm23[14],xmm18[15],xmm23[15]
-; AVX512BW-FCP-NEXT: vpshufb %xmm25, %xmm18, %xmm18
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm22[0,1,0,1],zmm18[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm22 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
-; AVX512BW-FCP-NEXT: vpermw %zmm0, %zmm22, %zmm22
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm29 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm3
+; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm28
+; AVX512BW-FCP-NEXT: vporq %xmm3, %xmm28, %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: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
+; 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, %xmm6, %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, %xmm7, %xmm28
+; AVX512BW-FCP-NEXT: vporq %xmm3, %xmm28, %xmm3
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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]
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm3[0,1,0,1],zmm6[0,1,0,1]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm28 {%k2}
+; 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 {{.*#+}} xmm6 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm4, %xmm7
+; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm7, %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: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,0,1],zmm4[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
+; AVX512BW-FCP-NEXT: vpermw %zmm17, %zmm4, %zmm4
; AVX512BW-FCP-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm18 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1}
; AVX512BW-FCP-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm5 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm12, %xmm6
-; AVX512BW-FCP-NEXT: vpshufb %xmm17, %xmm11, %xmm17
-; AVX512BW-FCP-NEXT: vporq %xmm6, %xmm17, %xmm6
-; 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: vpshufb %xmm9, %xmm11, %xmm9
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,0,1],zmm6[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm14, %xmm9
-; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm13, %xmm11
-; AVX512BW-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3],xmm13[4],xmm14[4],xmm13[5],xmm14[5],xmm13[6],xmm14[6],xmm13[7],xmm14[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm11, %xmm8
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm9[0,1,0,1]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm28 {%k1}
+; AVX512BW-FCP-NEXT: vpshufb %xmm9, %xmm22, %xmm3
+; AVX512BW-FCP-NEXT: vpshufb %xmm8, %xmm21, %xmm4
+; AVX512BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vpshufb %xmm12, %xmm4, %xmm4
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm4[0,1,0,1],zmm3[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpshufb %xmm2, %xmm24, %xmm2
+; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm23, %xmm0
+; AVX512BW-FCP-NEXT: vpor %xmm2, %xmm0, %xmm0
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm23[0],xmm24[0],xmm23[1],xmm24[1],xmm23[2],xmm24[2],xmm23[3],xmm24[3],xmm23[4],xmm24[4],xmm23[5],xmm24[5],xmm23[6],xmm24[6],xmm23[7],xmm24[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm31, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,0,1],zmm0[0,1,0,1]
; AVX512BW-FCP-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm8 {%k1}
-; AVX512BW-FCP-NEXT: vpshufb %xmm19, %xmm15, %xmm6
-; AVX512BW-FCP-NEXT: vpshufb %xmm24, %xmm16, %xmm9
-; AVX512BW-FCP-NEXT: vpor %xmm6, %xmm9, %xmm6
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
-; AVX512BW-FCP-NEXT: vpshufb %xmm4, %xmm9, %xmm4
-; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,0,1],zmm6[0,1,0,1]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
-; AVX512BW-FCP-NEXT: vpermw %zmm0, %zmm6, %zmm0
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
+; AVX512BW-FCP-NEXT: vpshufb %xmm1, %xmm25, %xmm1
+; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm27, %xmm2
+; AVX512BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm27[0],xmm25[0],xmm27[1],xmm25[1],xmm27[2],xmm25[2],xmm27[3],xmm25[3],xmm27[4],xmm25[4],xmm27[5],xmm25[5],xmm27[6],xmm25[6],xmm27[7],xmm25[7]
+; AVX512BW-FCP-NEXT: vpshufb %xmm30, %xmm2, %xmm2
+; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,0,1],zmm1[0,1,0,1]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512BW-FCP-NEXT: vpermw %zmm17, %zmm2, %zmm2
; AVX512BW-FCP-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm4 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k1}
; AVX512BW-FCP-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
; AVX512BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm8 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm1, %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: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm2, %zmm2
+; 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[21],zero,zmm2[19],zero,zero,zero,zero,zmm2[22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[55],zero,zero,zero,zero,zmm2[58],zero,zmm2[56],zero,zero,zero,zero,zmm2[59],zero,zmm2[57]
+; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[18],zero,zmm2[18,19,20,21],zero,zmm2[19],zero,zmm2[25,26,27,22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[56,57],zero,zmm2[55],zero,zmm2[53,54,55,58],zero,zmm2[56],zero,zmm2[60,61,58,59]
+; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero
+; 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 {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k2}
+; AVX512BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
+; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm1, %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, %ymm18, %zmm26, %zmm3
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm3[20],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero
+; AVX512BW-FCP-NEXT: vporq %zmm1, %zmm3, %zmm1
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512BW-FCP-NEXT: vpermw %zmm17, %zmm3, %zmm3
+; 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 %zmm3, %zmm1 {%k1}
+; AVX512BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 128(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 64(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 256(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 192(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 384(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rax)
+; AVX512BW-FCP-NEXT: addq $200, %rsp
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
; AVX512DQ-BW-LABEL: store_i8_stride7_vf64:
; AVX512DQ-BW: # %bb.0:
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r10), %zmm0
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %ymm13
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm13[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-NEXT: vpbroadcastd {{.*#+}} ymm3 = [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: movl $338170920, %r11d # imm = 0x14281428
-; AVX512DQ-BW-NEXT: kmovd %r11d, %k2
-; AVX512DQ-BW-NEXT: vpshufb %ymm3, %ymm2, %ymm1 {%k2}
-; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm2, %ymm4
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm9 = [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 %ymm9, %ymm13, %ymm5
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %ymm16
+; 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, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %ymm17
+; AVX512DQ-BW-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-BW-NEXT: vpshufb %ymm1, %ymm17, %ymm4
+; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm4, %ymm2
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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 %ymm4, %ymm16, %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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 %ymm5, %ymm17, %ymm5
; AVX512DQ-BW-NEXT: vpor %ymm4, %ymm5, %ymm4
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[2,3,2,3],zmm1[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm4 = [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 %ymm4, %ymm14, %ymm5
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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-BW-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpshufb %ymm10, %ymm15, %ymm6
-; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm6, %ymm6
-; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [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 %ymm5, %ymm15, %ymm7
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT: movl $676341840, %r11d # imm = 0x28502850
-; AVX512DQ-BW-NEXT: kmovd %r11d, %k3
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm11, %ymm7 {%k3}
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[2,3,2,3],zmm7[2,3,2,3]
-; AVX512DQ-BW-NEXT: movabsq $1742999440035548184, %r11 # imm = 0x183060C183060C18
-; AVX512DQ-BW-NEXT: kmovq %r11, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm1 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %ymm16
-; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128]
-; AVX512DQ-BW-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm16, %ymm11
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %ymm19
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
-; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm19, %ymm12
-; AVX512DQ-BW-NEXT: vpor %ymm11, %ymm12, %ymm11
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [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 %ymm17, %ymm19, %ymm12
-; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm18 = [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-BW-NEXT: # ymm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm18, %ymm16, %ymm20
-; AVX512DQ-BW-NEXT: vporq %ymm12, %ymm20, %ymm12
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm12[2,3,2,3],zmm11[2,3,2,3]
-; AVX512DQ-BW-NEXT: movabsq $6971997760142192736, %r11 # imm = 0x60C183060C183060
-; AVX512DQ-BW-NEXT: kmovq %r11, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm1 {%k1}
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm11 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm11, %zmm11
-; AVX512DQ-BW-NEXT: movabsq $-9150747060186627967, %r11 # imm = 0x8102040810204081
-; AVX512DQ-BW-NEXT: kmovq %r11, %k4
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm1 {%k4}
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm14, %ymm20
-; AVX512DQ-BW-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]
-; AVX512DQ-BW-NEXT: vpshufb %ymm12, %ymm15, %ymm21
-; AVX512DQ-BW-NEXT: vporq %ymm20, %ymm21, %ymm20
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm22 = [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]
-; AVX512DQ-BW-NEXT: vpshufb %ymm22, %ymm14, %ymm14
-; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm23 = [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-BW-NEXT: # ymm23 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm15, %ymm15
-; AVX512DQ-BW-NEXT: vpor %ymm14, %ymm15, %ymm14
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm14, %zmm20, %zmm14
-; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm13[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: 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: vpshufb %ymm25, %ymm2, %ymm15 {%k3}
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm13, %ymm13
-; AVX512DQ-BW-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-NEXT: vpshufb %ymm21, %ymm2, %ymm2
-; AVX512DQ-BW-NEXT: vpor %ymm2, %ymm13, %ymm2
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm15, %zmm2, %zmm2
-; AVX512DQ-BW-NEXT: movabsq $3485998880071096368, %r11 # imm = 0x3060C183060C1830
-; AVX512DQ-BW-NEXT: kmovq %r11, %k4
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm14, %zmm2 {%k4}
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm16, %ymm15
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm14 = [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 %ymm14, %ymm19, %ymm24
-; AVX512DQ-BW-NEXT: vporq %ymm15, %ymm24, %ymm15
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [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]
-; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm19, %ymm19
-; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm28 = [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-BW-NEXT: # ymm28 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm16, %ymm16
-; AVX512DQ-BW-NEXT: vporq %ymm19, %ymm16, %ymm16
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,3,2,3]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r10), %ymm16
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm29 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
-; AVX512DQ-BW-NEXT: vpermw %ymm16, %ymm29, %ymm19
-; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm16 = ymm16[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: vinserti64x4 $1, %ymm19, %zmm16, %zmm16
-; AVX512DQ-BW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
-; AVX512DQ-BW-NEXT: kmovq %r10, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm16, %zmm15 {%k5}
-; AVX512DQ-BW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
-; AVX512DQ-BW-NEXT: kmovq %r10, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm15, %zmm2 {%k5}
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm15
-; AVX512DQ-BW-NEXT: vpshufb %ymm4, %ymm15, %ymm4
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %ymm16
-; AVX512DQ-BW-NEXT: vpshufb %ymm10, %ymm16, %ymm10
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %ymm24
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %ymm23
+; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm23[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-NEXT: movl $676341840, %r10d # imm = 0x28502850
+; AVX512DQ-BW-NEXT: kmovd %r10d, %k1
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm4 {%k1} = ymm24[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm7 = [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 %ymm7, %ymm23, %ymm5
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm8 = [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 %ymm8, %ymm24, %ymm6
+; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm6, %ymm5
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm5
+; AVX512DQ-BW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
+; AVX512DQ-BW-NEXT: kmovq %r10, %k2
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm2, %zmm5 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm2 = [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 %ymm2, %ymm4, %ymm9
+; AVX512DQ-BW-NEXT: vmovdqa 32(%r8), %ymm10
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm6 = [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 %ymm6, %ymm10, %ymm11
+; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm11, %ymm9
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [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 %ymm11, %ymm10, %ymm10
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [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 %ymm11, %ymm4, %ymm4
; AVX512DQ-BW-NEXT: vpor %ymm4, %ymm10, %ymm4
-; AVX512DQ-BW-NEXT: vpshufb %ymm22, %ymm15, %ymm10
-; AVX512DQ-BW-NEXT: vpshufb %ymm23, %ymm16, %ymm19
-; AVX512DQ-BW-NEXT: vporq %ymm10, %ymm19, %ymm10
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm10[2,3,2,3],zmm4[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %ymm24
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %ymm27
-; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm27[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,1,1,4,4,5,5]
-; AVX512DQ-BW-NEXT: vpshufb %ymm25, %ymm24, %ymm10 {%k3}
-; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm24, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %ymm9, %ymm27, %ymm9
-; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm10[2,3,2,3],zmm8[2,3,2,3]
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm4 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %ymm19
-; AVX512DQ-BW-NEXT: vpshufb %ymm17, %ymm19, %ymm8
-; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %ymm22
-; AVX512DQ-BW-NEXT: vpshufb %ymm18, %ymm22, %ymm9
-; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %ymm26, %ymm19, %ymm9
-; AVX512DQ-BW-NEXT: vpshufb %ymm28, %ymm22, %ymm10
-; AVX512DQ-BW-NEXT: vpor %ymm9, %ymm10, %ymm9
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm9[2,3,2,3],zmm8[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm29, %zmm9
-; AVX512DQ-BW-NEXT: movabsq $1161999626690365456, %r10 # imm = 0x1020408102040810
-; AVX512DQ-BW-NEXT: kmovq %r10, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm8 {%k5}
-; AVX512DQ-BW-NEXT: movabsq $2033499346708139548, %r10 # imm = 0x1C3870E1C3870E1C
-; AVX512DQ-BW-NEXT: kmovq %r10, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm8, %zmm4 {%k5}
-; AVX512DQ-BW-NEXT: vpshufb %ymm11, %ymm15, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %ymm12, %ymm16, %ymm9
-; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm9, %ymm8
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %xmm17
-; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm10
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %xmm18
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm23 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm11, %xmm11
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm26
-; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm27, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %ymm21, %ymm24, %ymm11
-; AVX512DQ-BW-NEXT: vpor %ymm8, %ymm11, %ymm8
-; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm11
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %xmm20
-; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm12
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %xmm21
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm28 = 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-BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm28, %xmm28
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm28, %zmm8
-; AVX512DQ-BW-NEXT: movabsq $435749860008887046, %rcx # imm = 0x60C183060C18306
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm26, %zmm8 {%k5}
-; AVX512DQ-BW-NEXT: vpshufb %ymm13, %ymm22, %ymm13
-; AVX512DQ-BW-NEXT: vpshufb %ymm14, %ymm19, %ymm14
-; AVX512DQ-BW-NEXT: vporq %ymm13, %ymm14, %ymm30
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm13
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm26
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm14
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm9, %zmm4
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rax), %ymm9
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-NEXT: vpermw %ymm9, %ymm10, %ymm10
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[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: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-BW-NEXT: movabsq $145249953336295682, %rax # imm = 0x204081020408102
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm4 {%k3}
+; AVX512DQ-BW-NEXT: movabsq $-4357498600088870461, %rax # imm = 0xC3870E1C3870E1C3
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm4, %zmm5 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm14
+; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm14, %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm15
+; AVX512DQ-BW-NEXT: vpshufb %ymm8, %ymm15, %ymm7
+; AVX512DQ-BW-NEXT: vpor %ymm4, %ymm7, %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %xmm25
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %xmm26
+; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm26[8],xmm25[8],xmm26[9],xmm25[9],xmm26[10],xmm25[10],xmm26[11],xmm25[11],xmm26[12],xmm25[12],xmm26[13],xmm25[13],xmm26[14],xmm25[14],xmm26[15],xmm25[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm18, %ymm0
+; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT: vpshufb %ymm1, %ymm19, %ymm1
+; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm18[8],xmm19[9],xmm18[9],xmm19[10],xmm18[10],xmm19[11],xmm18[11],xmm19[12],xmm18[12],xmm19[13],xmm18[13],xmm19[14],xmm18[14],xmm19[15],xmm18[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-BW-NEXT: movabsq $435749860008887046, %rax # imm = 0x60C183060C18306
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %ymm20
+; AVX512DQ-BW-NEXT: vpshufb %ymm2, %ymm20, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %ymm21
+; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm21, %ymm1
+; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %xmm30
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm30[8],xmm2[9],xmm30[9],xmm2[10],xmm30[10],xmm2[11],xmm30[11],xmm2[12],xmm30[12],xmm2[13],xmm30[13],xmm2[14],xmm30[14],xmm2[15],xmm30[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: movabsq $2323999253380730912, %rax # imm = 0x2040810204081020
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3}
+; AVX512DQ-BW-NEXT: movabsq $4066998693416279096, %rax # imm = 0x3870E1C3870E1C38
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm4 {%k3}
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm14[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT: vpbroadcastd {{.*#+}} ymm7 = [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: movl $338170920, %eax # imm = 0x14281428
+; AVX512DQ-BW-NEXT: kmovd %eax, %k3
+; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm15, %ymm0 {%k3}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3],xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm6[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm18[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,2,3,3,4,6,7,7]
+; AVX512DQ-BW-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-BW-NEXT: vpshufb %ymm1, %ymm19, %ymm9
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm6, %ymm9 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %xmm29
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %xmm31
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm31[0],xmm29[0],xmm31[1],xmm29[1],xmm31[2],xmm29[2],xmm31[3],xmm29[3],xmm31[4],xmm29[4],xmm31[5],xmm29[5],xmm31[6],xmm29[6],xmm31[7],xmm29[7]
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm6, %xmm6
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm9[2,3,2,3],zmm6[0,1,0,1]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm22 {%k2}
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm20[27],zero,zero,zero,zero,ymm20[30],zero,ymm20[28],zero,zero,zero,zero,ymm20[31],zero
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm21[27],zero,zero,zero,zero,ymm21[30],zero,ymm21[28],zero,zero,zero,zero,ymm21[31],zero,ymm21[29]
+; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm6, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm27
; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm28
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm31 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm29 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm31, %xmm31
-; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,1,0,1]
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm30, %zmm31, %zmm30
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm31 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm31, %zmm31
-; AVX512DQ-BW-NEXT: movabsq $2323999253380730912, %rcx # imm = 0x2040810204081020
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm31, %zmm30 {%k5}
-; AVX512DQ-BW-NEXT: movabsq $4066998693416279096, %rcx # imm = 0x3870E1C3870E1C38
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k5
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm30, %zmm8 {%k5}
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm27 = ymm27[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm27 = ymm27[2,2,3,3,6,6,7,7]
-; AVX512DQ-BW-NEXT: vpshufb %ymm3, %ymm24, %ymm27 {%k2}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm20[0],xmm21[0],xmm20[1],xmm21[1],xmm20[2],xmm21[2],xmm20[3],xmm21[3],xmm20[4],xmm21[4],xmm20[5],xmm21[5],xmm20[6],xmm21[6],xmm20[7],xmm21[7]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm24 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm3, %xmm3
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm27[2,3,2,3],zmm3[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3],xmm17[4],xmm18[4],xmm17[5],xmm18[5],xmm17[6],xmm18[6],xmm17[7],xmm18[7]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm27 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm3, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %ymm5, %ymm16, %ymm5
-; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,2,3,3,4,6,7,7]
-; AVX512DQ-BW-NEXT: vmovdqu8 %ymm15, %ymm5 {%k3}
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm5[2,3,2,3],zmm3[0,1,0,1]
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm30, %zmm3 {%k4}
-; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm22, %ymm5
-; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm19, %ymm6
-; AVX512DQ-BW-NEXT: vpor %ymm5, %ymm6, %ymm6
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm28[0],xmm26[0],xmm28[1],xmm26[1],xmm28[2],xmm26[2],xmm28[3],xmm26[3],xmm28[4],xmm26[4],xmm28[5],xmm26[5],xmm28[6],xmm26[6],xmm28[7],xmm26[7]
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm5 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm7, %xmm7
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[2,3,2,3],zmm7[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [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 %zmm0, %zmm7, %zmm7
-; AVX512DQ-BW-NEXT: movabsq $580999813345182728, %rcx # imm = 0x810204081020408
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm6 {%k2}
-; AVX512DQ-BW-NEXT: movabsq $1016749673354069774, %rcx # imm = 0xE1C3870E1C3870E
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k2
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k2}
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm18, %xmm7
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm15 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm17, %xmm16
-; AVX512DQ-BW-NEXT: vporq %xmm7, %xmm16, %xmm7
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm16, %xmm16
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm7[0,1,0,1],zmm16[0,1,0,1]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm16 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-BW-NEXT: vpshufb %xmm16, %xmm21, %xmm17
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm18 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm18, %xmm20, %xmm19
-; AVX512DQ-BW-NEXT: vporq %xmm17, %xmm19, %xmm17
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm21[8],xmm20[8],xmm21[9],xmm20[9],xmm21[10],xmm20[10],xmm21[11],xmm20[11],xmm21[12],xmm20[12],xmm21[13],xmm20[13],xmm21[14],xmm20[14],xmm21[15],xmm20[15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm19, %xmm19
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm17[0,1,0,1],zmm19[0,1,0,1]
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm7, %zmm17 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm7 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm26, %xmm19
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm20 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm28, %xmm21
-; AVX512DQ-BW-NEXT: vporq %xmm19, %xmm21, %xmm19
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm28[8],xmm26[8],xmm28[9],xmm26[9],xmm28[10],xmm26[10],xmm28[11],xmm26[11],xmm28[12],xmm26[12],xmm28[13],xmm26[13],xmm28[14],xmm26[14],xmm28[15],xmm26[15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm21, %xmm21
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[0,1,0,1],zmm21[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm21 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm21, %zmm21
-; AVX512DQ-BW-NEXT: movabsq $290499906672591364, %rcx # imm = 0x408102040810204
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm21, %zmm19 {%k1}
-; AVX512DQ-BW-NEXT: movabsq $-8714997200177740921, %rcx # imm = 0x870E1C3870E1C387
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm19, %zmm17 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm10, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm9, %xmm15
-; AVX512DQ-BW-NEXT: vpor %xmm6, %xmm15, %xmm6
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm9, %xmm9
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,0,1],zmm6[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpshufb %xmm16, %xmm12, %xmm9
-; AVX512DQ-BW-NEXT: vpshufb %xmm18, %xmm11, %xmm10
-; AVX512DQ-BW-NEXT: vpor %xmm9, %xmm10, %xmm9
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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-NEXT: vpshufb %xmm24, %xmm10, %xmm10
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[0,1,0,1],zmm9[0,1,0,1]
-; AVX512DQ-BW-NEXT: movabsq $871499720017774092, %rcx # imm = 0xC183060C183060C
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm9 {%k1}
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm13, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm14, %xmm7
-; AVX512DQ-BW-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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: vpshufb %xmm5, %xmm7, %xmm5
-; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,0,1],zmm6[0,1,0,1]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm6, %zmm0
-; AVX512DQ-BW-NEXT: movabsq $4647998506761461824, %rcx # imm = 0x4081020408102040
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm5 {%k1}
-; AVX512DQ-BW-NEXT: movabsq $8133997386832558192, %rcx # imm = 0x70E1C3870E1C3870
-; AVX512DQ-BW-NEXT: kmovq %rcx, %k1
-; AVX512DQ-BW-NEXT: vmovdqu8 %zmm5, %zmm9 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 256(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm28[0],xmm27[0],xmm28[1],xmm27[1],xmm28[2],xmm27[2],xmm28[3],xmm27[3],xmm28[4],xmm27[4],xmm28[5],xmm27[5],xmm28[6],xmm27[6],xmm28[7],xmm27[7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm6[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [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 %zmm3, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512DQ-BW-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512DQ-BW-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm22 {%k2}
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm23[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
+; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm24, %ymm0 {%k3}
+; AVX512DQ-BW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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: # ymm7 = mem[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm24, %ymm6
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [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 %ymm24, %ymm23, %ymm23
+; AVX512DQ-BW-NEXT: vporq %ymm6, %ymm23, %ymm6
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm6[2,3,2,3],zmm0[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[24,25],zero,ymm16[23],zero,ymm16[21,22,23,26],zero,ymm16[24],zero,ymm16[28,29,26,27]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm6 = ymm17[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm17[25],zero,ymm17[23],zero,zero,zero,zero,ymm17[26],zero,ymm17[24],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm6, %ymm0
+; AVX512DQ-BW-NEXT: vpshufb %ymm1, %ymm17, %ymm1
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,2,3,3,4,6,7,7]
+; AVX512DQ-BW-NEXT: vmovdqu8 %ymm6, %ymm1 {%k1}
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[2,3,2,3],zmm1[2,3,2,3]
+; AVX512DQ-BW-NEXT: movabsq $1742999440035548184, %rax # imm = 0x183060C183060C18
+; AVX512DQ-BW-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm23 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm0
+; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm1
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[4,5,6,7],zmm1[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm6 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm6[23],zero,zmm6[23,24,25,26],zero,zmm6[24],zero,zmm6[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm6[59],zero,zero,zero,zero,zmm6[62],zero,zmm6[60],zero,zero,zero,zero,zmm6[63],zero
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[4,5,6,7],zmm0[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm16 = zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm16[25],zero,zmm16[23],zero,zero,zero,zero,zmm16[26],zero,zmm16[24],zero,zero,zmm16[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm16[59],zero,zero,zero,zero,zmm16[62],zero,zmm16[60],zero,zero,zero,zero,zmm16[63],zero,zmm16[61]
+; AVX512DQ-BW-NEXT: vporq %zmm6, %zmm16, %zmm6
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512DQ-BW-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm23 {%k2}
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm6, %zmm6
+; AVX512DQ-BW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm6, %zmm23 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm17 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm26, %xmm16
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm6 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm25, %xmm13
+; AVX512DQ-BW-NEXT: vporq %xmm16, %xmm13, %xmm13
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm25[0],xmm26[0],xmm25[1],xmm26[1],xmm25[2],xmm26[2],xmm25[3],xmm26[3],xmm25[4],xmm26[4],xmm25[5],xmm26[5],xmm25[6],xmm26[6],xmm25[7],xmm26[7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm16, %xmm12
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,0,1],zmm13[0,1,0,1]
+; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm19, %xmm12
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm13 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %xmm13, %xmm18, %xmm26
+; AVX512DQ-BW-NEXT: vporq %xmm12, %xmm26, %xmm12
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm18[0],xmm19[0],xmm18[1],xmm19[1],xmm18[2],xmm19[2],xmm18[3],xmm19[3],xmm18[4],xmm19[4],xmm18[5],xmm19[5],xmm18[6],xmm19[6],xmm18[7],xmm19[7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm26, %xmm11
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm11[0,1,0,1],zmm12[0,1,0,1]
+; AVX512DQ-BW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm11, %zmm16 {%k3}
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm11 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm30, %xmm26
+; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} xmm12 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm2, %xmm9
+; AVX512DQ-BW-NEXT: vporq %xmm26, %xmm9, %xmm9
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3],xmm2[4],xmm30[4],xmm2[5],xmm30[5],xmm2[6],xmm30[6],xmm2[7],xmm30[7]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,0,1],zmm9[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm9, %zmm9
+; AVX512DQ-BW-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm2 {%k3}
+; AVX512DQ-BW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
+; AVX512DQ-BW-NEXT: kmovq %rax, %k3
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm2, %zmm16 {%k3}
+; AVX512DQ-BW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm14[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,1,1,4,4,5,5]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm2 {%k1} = ymm15[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
+; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm15, %ymm7
+; AVX512DQ-BW-NEXT: vpshufb %ymm24, %ymm14, %ymm9
+; AVX512DQ-BW-NEXT: vpor %ymm7, %ymm9, %ymm7
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[2,3,2,3],zmm7[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm7 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm18[18],zero,zmm18[18,19,20,21],zero,zmm18[19],zero,zmm18[25,26,27,22],zero,zmm18[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm18[56,57],zero,zmm18[55],zero,zmm18[53,54,55,58],zero,zmm18[56],zero,zmm18[60,61,58,59]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm9 = zmm19[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm19[18],zero,zero,zero,zero,zmm19[21],zero,zmm19[19],zero,zero,zero,zero,zmm19[22],zero,zmm19[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm19[57],zero,zmm19[55],zero,zero,zero,zero,zmm19[58],zero,zmm19[56],zero,zero,zero,zero
+; AVX512DQ-BW-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm2, %zmm7 {%k2}
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[20],zero,zmm0[18],zero,zmm0[20,21,20,21],zero,zmm0[19],zero,zmm0[19,20,21,22],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm0[56,57,56,57],zero,zmm0[55],zero,zmm0[55,56,57,58],zero,zmm0[56],zero,zmm0[62,63]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm1[20],zero,zmm1[18],zero,zero,zero,zero,zmm1[21],zero,zmm1[19],zero,zero,zero,zero,zmm1[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm1[57],zero,zmm1[55],zero,zero,zero,zero,zmm1[58],zero,zmm1[56],zero,zero
+; AVX512DQ-BW-NEXT: vporq %zmm0, %zmm1, %zmm0
+; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm1 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
+; AVX512DQ-BW-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512DQ-BW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512DQ-BW-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm7 {%k1}
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm29, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm13, %xmm31, %xmm1
+; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm29[8],xmm31[8],xmm29[9],xmm31[9],xmm29[10],xmm31[10],xmm29[11],xmm31[11],xmm29[12],xmm31[12],xmm29[13],xmm31[13],xmm29[14],xmm31[14],xmm29[15],xmm31[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,0,1],zmm1[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm8, %xmm2
+; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,0,1],zmm2[0,1,0,1]
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
+; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm27, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm28, %xmm2
+; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm28[8],xmm27[8],xmm28[9],xmm27[9],xmm28[10],xmm27[10],xmm28[11],xmm27[11],xmm28[12],xmm27[12],xmm28[13],xmm27[13],xmm28[14],xmm27[14],xmm28[15],xmm27[15]
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,0,1],zmm2[0,1,0,1]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm2 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
+; AVX512DQ-BW-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm2, %zmm0 {%k1}
+; AVX512DQ-BW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
+; AVX512DQ-BW-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
+; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 256(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, 384(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 192(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 64(%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 $200, %rsp
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128]
-; AVX512DQ-BW-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm5, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [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 %ymm7, %ymm13, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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-BW-FCP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm5, %ymm6, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[2,3,2,3],zmm1[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %ymm16
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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-BW-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm16, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm6, %ymm1
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm18 = [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-FCP-NEXT: # ymm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm18, %ymm16, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [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 %ymm19, %ymm15, %ymm11
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm8
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm7
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm17
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %ymm3
+; AVX512DQ-BW-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-BW-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
+; AVX512DQ-BW-FCP-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-BW-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm11
; AVX512DQ-BW-FCP-NEXT: vpor %ymm6, %ymm11, %ymm6
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm6[2,3,2,3],zmm1[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm1 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %ymm17
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128]
-; AVX512DQ-BW-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm17, %ymm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %ymm22
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm22, %ymm12
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm11, %ymm12, %ymm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm22, %ymm12
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm21 = [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-BW-FCP-NEXT: # ymm21 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm21, %ymm17, %ymm14
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm12, %ymm14, %ymm12
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm12[2,3,2,3],zmm11[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: movabsq $6971997760142192736, %r10 # imm = 0x60C183060C183060
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm1 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm11 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm0, %zmm11, %zmm11
-; AVX512DQ-BW-FCP-NEXT: movabsq $-9150747060186627967, %r10 # imm = 0x8102040810204081
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm1 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [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 %ymm11, %ymm13, %ymm14
-; AVX512DQ-BW-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]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm14, %ymm23, %ymm14
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm24 = [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]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm24, %ymm13, %ymm13
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm25 = [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-BW-FCP-NEXT: # ymm25 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm25, %ymm2, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm13, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm23
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm15, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [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 %ymm14, %ymm16, %ymm26
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm2, %ymm26, %ymm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm26 = [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]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm15, %ymm15
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm27 = [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-BW-FCP-NEXT: # ymm27 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm27, %ymm16, %ymm16
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm15, %ymm16, %ymm15
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm2, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm3, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm10, %ymm10
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm10, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm11
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %ymm6
+; AVX512DQ-BW-FCP-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-BW-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm10
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm12
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [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 %ymm13, %ymm12, %ymm14
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm10, %ymm14, %ymm14
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm6, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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 %ymm9, %ymm12, %ymm12
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm6, %ymm12, %ymm6
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm14, %zmm10
; AVX512DQ-BW-FCP-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
-; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm23, %zmm2 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [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 %ymm15, %ymm17, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm16 = [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 %ymm16, %ymm22, %ymm28
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm23, %ymm28, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm28 = [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]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm28, %ymm22, %ymm22
-; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm29 = [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-BW-FCP-NEXT: # ymm29 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm29, %ymm17, %ymm17
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm22, %ymm17, %ymm17
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm23, %zmm17
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rax), %ymm22
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm30 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
-; AVX512DQ-BW-FCP-NEXT: vpermw %ymm22, %ymm30, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm22 = ymm22[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: vinserti64x4 $1, %ymm23, %zmm22, %zmm22
+; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm11, %zmm10 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm14
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm19 = [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 %ymm19, %ymm14, %ymm11
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %ymm12
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} ymm20 = [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 %ymm20, %ymm12, %ymm15
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm11, %ymm15, %ymm15
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [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 %ymm6, %ymm12, %ymm16
+; 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, %ymm14, %ymm14
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm16, %ymm14, %ymm14
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm15, %zmm14
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm16 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512DQ-BW-FCP-NEXT: vpermw %ymm15, %ymm16, %ymm16
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[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: vinserti64x4 $1, %ymm16, %zmm15, %zmm15
; AVX512DQ-BW-FCP-NEXT: movabsq $145249953336295682, %rax # imm = 0x204081020408102
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm17 {%k3}
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm15, %zmm14 {%k2}
; AVX512DQ-BW-FCP-NEXT: movabsq $-4357498600088870461, %rax # imm = 0xC3870E1C3870E1C3
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm2 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm17
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm18, %ymm17, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm22, %ymm19
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm18, %ymm19, %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm26, %ymm22, %ymm19
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm27, %ymm17, %ymm23
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm19, %ymm23, %ymm19
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm19[2,3,2,3],zmm18[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm18
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm18, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %ymm23
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm10, %ymm23, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm7, %ymm10, %ymm7
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm24, %ymm18, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm25, %ymm23, %ymm24
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm10, %ymm24, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm10[2,3,2,3],zmm7[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm7 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %ymm19
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm19, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %ymm20
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm21, %ymm20, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm10, %ymm21, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm28, %ymm19, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm29, %ymm20, %ymm24
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm21, %ymm24, %ymm21
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm21[2,3,2,3],zmm10[2,3,2,3]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm0, %zmm30, %zmm21
-; AVX512DQ-BW-FCP-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm21, %zmm10 {%k3}
-; AVX512DQ-BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm7 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm11, %ymm18, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm12, %ymm23, %ymm11
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm10, %ymm11, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = 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-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm24, %xmm24
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm24[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm24, %zmm25
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm13, %ymm22, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm14, %ymm17, %ymm13
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm10, %ymm13, %ymm10
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm13
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm26 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm24, %xmm26, %xmm26
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm26[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm26, %zmm10
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm10 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm14
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm21
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %xmm22
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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: vpshufb {{.*#+}} xmm2 = xmm2[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm16
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm16, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm11
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm13, %ymm11, %ymm3
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm23
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm24
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm24[8],xmm23[8],xmm24[9],xmm23[9],xmm24[10],xmm23[10],xmm24[11],xmm23[11],xmm24[12],xmm23[12],xmm24[13],xmm23[13],xmm24[14],xmm23[14],xmm24[15],xmm23[15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm13
; AVX512DQ-BW-FCP-NEXT: movabsq $435749860008887046, %rax # imm = 0x60C183060C18306
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm25, %zmm10 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm15, %ymm20, %ymm15
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm16, %ymm19, %ymm16
-; AVX512DQ-BW-FCP-NEXT: vporq %ymm15, %ymm16, %ymm26
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm15
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %xmm16
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm27 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm25 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm25, %xmm27, %xmm27
-; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm27 = ymm27[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm27, %zmm26
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm27 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm0, %zmm27, %zmm27
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm13 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %ymm18
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm18, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %ymm19
+; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm19, %ymm2
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %xmm25
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %xmm27
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm27[8],xmm25[8],xmm27[9],xmm25[9],xmm27[10],xmm25[10],xmm27[11],xmm25[11],xmm27[12],xmm25[12],xmm27[13],xmm25[13],xmm27[14],xmm25[14],xmm27[15],xmm25[15]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6,6,1,6,0,7,1,0,7,14,9,14,8,15,9,8,15]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm17, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT: movabsq $2323999253380730912, %rax # imm = 0x2040810204081020
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm27, %zmm26 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm27
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k2}
; AVX512DQ-BW-FCP-NEXT: movabsq $4066998693416279096, %rax # imm = 0x3870E1C3870E1C38
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm13 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm5[4,5,6,7],zmm4[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: vshufi64x2 {{.*#+}} zmm2 = zmm4[4,5,6,7],zmm5[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: vporq %zmm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm7[4,5,6,7],zmm8[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: vshufi64x2 {{.*#+}} zmm3 = zmm8[4,5,6,7],zmm7[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: vporq %zmm2, %zmm3, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm26
+; 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, %rax # imm = 0x183060C183060C18
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm0[4,5,6,7],zmm26[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: vshufi64x2 {{.*#+}} zmm2 = zmm26[4,5,6,7],zmm0[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: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm1 = [27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28,30,29,30,31,31,30,30,31,30,29,30,31,31,30,30,31]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm17, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k3
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm26, %zmm10 {%k3}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm26
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm8, %ymm22, %ymm8
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm22
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm17, %ymm9
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm8, %ymm9, %ymm9
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm26[0],xmm22[0],xmm26[1],xmm22[1],xmm26[2],xmm22[2],xmm26[3],xmm22[3],xmm26[4],xmm22[4],xmm26[5],xmm22[5],xmm26[6],xmm22[6],xmm26[7],xmm22[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm17, %xmm17
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm9[2,3,2,3],zmm17[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm28
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm3, %ymm23, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm23
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm18, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm27[0],xmm28[0],xmm27[1],xmm28[1],xmm27[2],xmm28[2],xmm27[3],xmm28[3],xmm27[4],xmm28[4],xmm27[5],xmm28[5],xmm27[6],xmm28[6],xmm27[7],xmm28[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm9, %xmm4, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[2,3,2,3],zmm4[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm18
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm17, %zmm3 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm20, %ymm4
-; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm19, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vpor %ymm4, %ymm5, %ymm5
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm18[0],xmm23[0],xmm18[1],xmm23[1],xmm18[2],xmm23[2],xmm18[3],xmm23[3],xmm18[4],xmm23[4],xmm18[5],xmm23[5],xmm18[6],xmm23[6],xmm18[7],xmm23[7]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm4, %xmm6, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[2,3,2,3],zmm6[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [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 %zmm0, %zmm6, %zmm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm20 {%k3}
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm16[28,29,30],zero,ymm16[28],zero,ymm16[30,31,30,31],zero,ymm16[29],zero,ymm16[31,28,29]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29],zero,zero,zero
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm7
+; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm6
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm31 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm31, %xmm3, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm1[2,3,2,3],zmm3[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27],zero,zero,zero,zero,ymm15[30],zero,ymm15[28],zero,zero,zero,zero,ymm15[31],zero,ymm15[29],zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm29 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vporq %ymm1, %ymm29, %ymm29
+; 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: vmovdqa {{.*#+}} xmm12 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm29[2,3,2,3],zmm0[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm29 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm18[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm18[27],zero,zero,zero,zero,ymm18[30],zero,ymm18[28],zero,zero,zero,zero,ymm18[31],zero
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm19[27],zero,zero,zero,zero,ymm19[30],zero,ymm19[28],zero,zero,zero,zero,ymm19[31],zero,ymm19[29]
+; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm3, %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 {{.*#+}} xmm30 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm30, %xmm0, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[2,3,2,3],zmm0[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 %zmm17, %zmm3, %zmm3
; AVX512DQ-BW-FCP-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm5 {%k2}
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k2
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm5, %zmm3 {%k2}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm28, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm17 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm27, %xmm19
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm5, %xmm19, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm28[8],xmm27[8],xmm28[9],xmm27[9],xmm28[10],xmm27[10],xmm28[11],xmm27[11],xmm28[12],xmm27[12],xmm28[13],xmm27[13],xmm28[14],xmm27[14],xmm28[15],xmm27[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm19, %xmm19
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm5[0,1,0,1],zmm19[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm20 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm22, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm26, %xmm27
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm5, %xmm27, %xmm5
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm22[8],xmm26[8],xmm22[9],xmm26[9],xmm22[10],xmm26[10],xmm22[11],xmm26[11],xmm22[12],xmm26[12],xmm22[13],xmm26[13],xmm22[14],xmm26[14],xmm22[15],xmm26[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm24, %xmm22, %xmm22
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,0,1],zmm22[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm19, %zmm5 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm19 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm23, %xmm22
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm24, %xmm18, %xmm26
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm22, %xmm26, %xmm22
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm18[8],xmm23[8],xmm18[9],xmm23[9],xmm18[10],xmm23[10],xmm18[11],xmm23[11],xmm18[12],xmm23[12],xmm18[13],xmm23[13],xmm18[14],xmm23[14],xmm18[15],xmm23[15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm25, %xmm18, %xmm18
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm22[0,1,0,1],zmm18[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm22 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm0, %zmm22, %zmm22
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm29 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm28
+; AVX512DQ-BW-FCP-NEXT: vporq %xmm3, %xmm28, %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: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
+; 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, %xmm6, %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, %xmm7, %xmm28
+; AVX512DQ-BW-FCP-NEXT: vporq %xmm3, %xmm28, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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-BW-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm3[0,1,0,1],zmm6[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm28 {%k2}
+; 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 {{.*#+}} xmm6 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm4, %xmm7
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm7, %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: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,0,1],zmm4[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20,20,21,20,21,21,22,21,22,20,21,20,21,21,22,21,22]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm17, %zmm4, %zmm4
; AVX512DQ-BW-FCP-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm18 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm18, %zmm5 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm12, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm17, %xmm11, %xmm17
-; AVX512DQ-BW-FCP-NEXT: vporq %xmm6, %xmm17, %xmm6
-; 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: vpshufb %xmm9, %xmm11, %xmm9
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,0,1],zmm6[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm14, %xmm9
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm13, %xmm11
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3],xmm13[4],xmm14[4],xmm13[5],xmm14[5],xmm13[6],xmm14[6],xmm13[7],xmm14[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm11, %xmm8
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm9[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm28 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm9, %xmm22, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm8, %xmm21, %xmm4
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm3, %xmm4, %xmm3
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vpshufb %xmm12, %xmm4, %xmm4
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm4[0,1,0,1],zmm3[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm2, %xmm24, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm23, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm2, %xmm0, %xmm0
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm23[0],xmm24[0],xmm23[1],xmm24[1],xmm23[2],xmm24[2],xmm23[3],xmm24[3],xmm23[4],xmm24[4],xmm23[5],xmm24[5],xmm23[6],xmm24[6],xmm23[7],xmm24[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm31, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,0,1],zmm0[0,1,0,1]
; AVX512DQ-BW-FCP-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm6, %zmm8 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm19, %xmm15, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm24, %xmm16, %xmm9
-; AVX512DQ-BW-FCP-NEXT: vpor %xmm6, %xmm9, %xmm6
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm4, %xmm9, %xmm4
-; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,0,1],zmm6[0,1,0,1]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
-; AVX512DQ-BW-FCP-NEXT: vpermw %zmm0, %zmm6, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm1, %xmm25, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm27, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vpor %xmm1, %xmm2, %xmm1
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm27[0],xmm25[0],xmm27[1],xmm25[1],xmm27[2],xmm25[2],xmm27[3],xmm25[3],xmm27[4],xmm25[4],xmm27[5],xmm25[5],xmm27[6],xmm25[6],xmm27[7],xmm25[7]
+; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm30, %xmm2, %xmm2
+; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,0,1],zmm1[0,1,0,1]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm17, %zmm2, %zmm2
; AVX512DQ-BW-FCP-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm0, %zmm4 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm2, %zmm1 {%k1}
; AVX512DQ-BW-FCP-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm4, %zmm8 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm1, %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: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm2, %zmm2
+; 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[21],zero,zmm2[19],zero,zero,zero,zero,zmm2[22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[55],zero,zero,zero,zero,zmm2[58],zero,zmm2[56],zero,zero,zero,zero,zmm2[59],zero,zmm2[57]
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm2, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[18],zero,zmm2[18,19,20,21],zero,zmm2[19],zero,zmm2[25,26,27,22],zero,zmm2[20],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zmm2[56,57],zero,zmm2[55],zero,zmm2[53,54,55,58],zero,zmm2[56],zero,zmm2[60,61,58,59]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %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,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22],zero,zmm3[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero,zero,zero
+; 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 {{.*#+}} zmm2 = zmm2[2,3,2,3,6,7,6,7]
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k2}
+; AVX512DQ-BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm1 # 64-byte Reload
+; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm1, %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, %ymm18, %zmm26, %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,zmm3[20],zero,zmm3[18],zero,zero,zero,zero,zmm3[21],zero,zmm3[19],zero,zero,zero,zero,zmm3[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm3[57],zero,zmm3[55],zero,zero,zero,zero,zmm3[58],zero,zmm3[56],zero,zero
+; AVX512DQ-BW-FCP-NEXT: vporq %zmm1, %zmm3, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12]
+; AVX512DQ-BW-FCP-NEXT: vpermw %zmm17, %zmm3, %zmm3
+; 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 %zmm3, %zmm1 {%k1}
+; AVX512DQ-BW-FCP-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512DQ-BW-FCP-NEXT: kmovq %rax, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 256(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 192(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 384(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT: addq $200, %rsp
; 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-reduce-ctpop.ll b/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
index 4d12edde75176..29366f74da12a 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-ctpop.ll
@@ -1487,81 +1487,79 @@ define <8 x i32> @reduce_ctpop_v4i64_buildvector_v8i32(<4 x i64> %a0, <4 x i64>
; AVX512VL-NEXT: vpsrlw $4, %ymm0, %ymm0
; AVX512VL-NEXT: vpand %ymm0, %ymm8, %ymm0
; AVX512VL-NEXT: vpshufb %ymm0, %ymm10, %ymm0
-; AVX512VL-NEXT: vpaddb %ymm0, %ymm9, %ymm9
-; AVX512VL-NEXT: vpxor %xmm0, %xmm0, %xmm0
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm9, %ymm9
+; AVX512VL-NEXT: vpaddb %ymm0, %ymm9, %ymm0
+; AVX512VL-NEXT: vpxor %xmm9, %xmm9, %xmm9
+; AVX512VL-NEXT: vpsadbw %ymm0, %ymm9, %ymm0
; AVX512VL-NEXT: vpand %ymm1, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm1, %ymm1
; AVX512VL-NEXT: vpand %ymm1, %ymm8, %ymm1
; AVX512VL-NEXT: vpshufb %ymm1, %ymm10, %ymm1
; AVX512VL-NEXT: vpaddb %ymm1, %ymm11, %ymm1
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm1, %ymm1
+; AVX512VL-NEXT: vpsadbw %ymm1, %ymm9, %ymm1
; AVX512VL-NEXT: vpand %ymm2, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm2, %ymm2
; AVX512VL-NEXT: vpand %ymm2, %ymm8, %ymm2
; AVX512VL-NEXT: vpshufb %ymm2, %ymm10, %ymm2
; AVX512VL-NEXT: vpaddb %ymm2, %ymm11, %ymm2
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm2, %ymm2
+; AVX512VL-NEXT: vpsadbw %ymm2, %ymm9, %ymm2
; AVX512VL-NEXT: vpand %ymm3, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm3, %ymm3
; AVX512VL-NEXT: vpand %ymm3, %ymm8, %ymm3
; AVX512VL-NEXT: vpshufb %ymm3, %ymm10, %ymm3
; AVX512VL-NEXT: vpaddb %ymm3, %ymm11, %ymm3
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm3, %ymm3
+; AVX512VL-NEXT: vpsadbw %ymm3, %ymm9, %ymm3
; AVX512VL-NEXT: vpand %ymm4, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm4, %ymm4
; AVX512VL-NEXT: vpand %ymm4, %ymm8, %ymm4
; AVX512VL-NEXT: vpshufb %ymm4, %ymm10, %ymm4
; AVX512VL-NEXT: vpaddb %ymm4, %ymm11, %ymm4
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm4, %ymm4
+; AVX512VL-NEXT: vpsadbw %ymm4, %ymm9, %ymm4
; AVX512VL-NEXT: vpand %ymm5, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm5, %ymm5
; AVX512VL-NEXT: vpand %ymm5, %ymm8, %ymm5
; AVX512VL-NEXT: vpshufb %ymm5, %ymm10, %ymm5
; AVX512VL-NEXT: vpaddb %ymm5, %ymm11, %ymm5
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm5, %ymm5
+; AVX512VL-NEXT: vpsadbw %ymm5, %ymm9, %ymm5
; AVX512VL-NEXT: vpand %ymm6, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm6, %ymm6
; AVX512VL-NEXT: vpand %ymm6, %ymm8, %ymm6
; AVX512VL-NEXT: vpshufb %ymm6, %ymm10, %ymm6
; AVX512VL-NEXT: vpaddb %ymm6, %ymm11, %ymm6
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm6, %ymm6
+; AVX512VL-NEXT: vpsadbw %ymm6, %ymm9, %ymm6
; AVX512VL-NEXT: vpand %ymm7, %ymm8, %ymm11
; AVX512VL-NEXT: vpshufb %ymm11, %ymm10, %ymm11
; AVX512VL-NEXT: vpsrlw $4, %ymm7, %ymm7
; AVX512VL-NEXT: vpand %ymm7, %ymm8, %ymm7
; AVX512VL-NEXT: vpshufb %ymm7, %ymm10, %ymm7
; AVX512VL-NEXT: vpaddb %ymm7, %ymm11, %ymm7
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm7, %ymm7
-; AVX512VL-NEXT: vpmovqb %ymm9, %xmm8
-; AVX512VL-NEXT: vpxor %xmm9, %xmm9, %xmm9
-; AVX512VL-NEXT: vpsadbw %xmm9, %xmm8, %xmm8
+; AVX512VL-NEXT: vpsadbw %ymm7, %ymm9, %ymm7
+; AVX512VL-NEXT: vpmovqb %ymm0, %xmm0
; AVX512VL-NEXT: vpmovqb %ymm1, %xmm1
-; AVX512VL-NEXT: vpsadbw %xmm1, %xmm9, %xmm1
; AVX512VL-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512VL-NEXT: vpsadbw %xmm2, %xmm9, %xmm2
; AVX512VL-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512VL-NEXT: vpsadbw %xmm3, %xmm9, %xmm3
; AVX512VL-NEXT: vpmovqb %ymm4, %xmm4
; AVX512VL-NEXT: vpmovqb %ymm5, %xmm5
; AVX512VL-NEXT: vpmovqb %ymm6, %xmm6
; AVX512VL-NEXT: vpmovqb %ymm7, %xmm7
-; AVX512VL-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm6, %ymm6
; AVX512VL-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4
-; AVX512VL-NEXT: vpsadbw %ymm0, %ymm4, %ymm0
-; AVX512VL-NEXT: vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,0,4,8,12]
-; AVX512VL-NEXT: vpermi2d %ymm6, %ymm0, %ymm4
-; AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]
-; AVX512VL-NEXT: vpbroadcastq {{.*#+}} xmm1 = [0,8,0,8]
-; AVX512VL-NEXT: vpermi2d %ymm3, %ymm2, %ymm1
-; AVX512VL-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512VL-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm5
+; AVX512VL-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512VL-NEXT: vpxor %xmm5, %xmm5, %xmm5
+; AVX512VL-NEXT: vpsadbw %zmm5, %zmm4, %zmm4
+; AVX512VL-NEXT: vpmovsxbd {{.*#+}} ymm6 = [8,12,8,12,8,12,0,4]
+; AVX512VL-NEXT: vpermd %zmm4, %zmm6, %zmm4
+; AVX512VL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX512VL-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512VL-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VL-NEXT: vpsadbw %zmm5, %zmm0, %zmm0
+; AVX512VL-NEXT: vpmovsxbd {{.*#+}} xmm1 = [8,12,0,4]
+; AVX512VL-NEXT: vpermd %zmm0, %zmm1, %zmm0
; AVX512VL-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512VL-NEXT: retq
;
@@ -1576,30 +1574,27 @@ define <8 x i32> @reduce_ctpop_v4i64_buildvector_v8i32(<4 x i64> %a0, <4 x i64>
; AVX512VPOPCNT-NEXT: vpopcntq %ymm6, %ymm6
; AVX512VPOPCNT-NEXT: vpopcntq %ymm7, %ymm7
; AVX512VPOPCNT-NEXT: vpmovqb %ymm0, %xmm0
-; AVX512VPOPCNT-NEXT: vpxor %xmm8, %xmm8, %xmm8
-; AVX512VPOPCNT-NEXT: vpsadbw %xmm0, %xmm8, %xmm0
; AVX512VPOPCNT-NEXT: vpmovqb %ymm1, %xmm1
-; AVX512VPOPCNT-NEXT: vpsadbw %xmm1, %xmm8, %xmm1
; AVX512VPOPCNT-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512VPOPCNT-NEXT: vpsadbw %xmm2, %xmm8, %xmm2
; AVX512VPOPCNT-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512VPOPCNT-NEXT: vpsadbw %xmm3, %xmm8, %xmm3
; AVX512VPOPCNT-NEXT: vpmovqb %ymm4, %xmm4
; AVX512VPOPCNT-NEXT: vpmovqb %ymm5, %xmm5
; AVX512VPOPCNT-NEXT: vpmovqb %ymm6, %xmm6
; AVX512VPOPCNT-NEXT: vpmovqb %ymm7, %xmm7
-; AVX512VPOPCNT-NEXT: vpbroadcastq {{.*#+}} xmm8 = [0,8,0,8]
-; AVX512VPOPCNT-NEXT: vpermt2d %ymm3, %ymm8, %ymm2
-; AVX512VPOPCNT-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; AVX512VPOPCNT-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]
-; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm1
-; AVX512VPOPCNT-NEXT: vpxor %xmm2, %xmm2, %xmm2
-; AVX512VPOPCNT-NEXT: vpsadbw %ymm2, %ymm1, %ymm1
-; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm3
-; AVX512VPOPCNT-NEXT: vpsadbw %ymm2, %ymm3, %ymm2
-; AVX512VPOPCNT-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,0,0,0,0,4,8,12]
-; AVX512VPOPCNT-NEXT: vpermi2d %ymm1, %ymm2, %ymm3
-; AVX512VPOPCNT-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4
+; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm5
+; AVX512VPOPCNT-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512VPOPCNT-NEXT: vpxor %xmm5, %xmm5, %xmm5
+; AVX512VPOPCNT-NEXT: vpsadbw %zmm5, %zmm4, %zmm4
+; AVX512VPOPCNT-NEXT: vpmovsxbd {{.*#+}} ymm6 = [8,12,8,12,8,12,0,4]
+; AVX512VPOPCNT-NEXT: vpermd %zmm4, %zmm6, %zmm4
+; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX512VPOPCNT-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm1
+; AVX512VPOPCNT-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512VPOPCNT-NEXT: vpsadbw %zmm5, %zmm0, %zmm0
+; AVX512VPOPCNT-NEXT: vpmovsxbd {{.*#+}} xmm1 = [8,12,0,4]
+; AVX512VPOPCNT-NEXT: vpermd %zmm0, %zmm1, %zmm0
+; AVX512VPOPCNT-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512VPOPCNT-NEXT: retq
%p0 = tail call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %a0)
%p1 = tail call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %a1)
diff --git a/llvm/test/CodeGen/X86/widen_fdiv.ll b/llvm/test/CodeGen/X86/widen_fdiv.ll
index fd251a99ca500..2d9e3f60bee46 100644
--- a/llvm/test/CodeGen/X86/widen_fdiv.ll
+++ b/llvm/test/CodeGen/X86/widen_fdiv.ll
@@ -194,35 +194,21 @@ define void @widen_fdiv_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {
; AVX512VL-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
; AVX512VL-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero
; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm5 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm6 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm7 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm8 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm9 = mem[0],zero
-; AVX512VL-NEXT: vdivps %xmm9, %xmm8, %xmm8
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm9 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm10 = mem[0],zero
-; AVX512VL-NEXT: vdivps %xmm10, %xmm9, %xmm9
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm10 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm11 = mem[0],zero
-; AVX512VL-NEXT: vdivps %xmm11, %xmm10, %xmm10
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm11 = mem[0],zero
-; AVX512VL-NEXT: vmovsd {{.*#+}} xmm12 = mem[0],zero
-; AVX512VL-NEXT: vdivps %xmm12, %xmm11, %xmm11
-; AVX512VL-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6
-; AVX512VL-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
-; AVX512VL-NEXT: vdivps %ymm6, %ymm2, %ymm2
-; AVX512VL-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm3
-; AVX512VL-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512VL-NEXT: vdivps %ymm3, %ymm0, %ymm0
-; AVX512VL-NEXT: vpmovsxbq {{.*#+}} ymm1 = [0,2,4,6]
-; AVX512VL-NEXT: vpermi2pd %ymm2, %ymm0, %ymm1
-; AVX512VL-NEXT: vinsertf32x4 $1, %xmm11, %zmm10, %zmm0
-; AVX512VL-NEXT: vinsertf32x4 $1, %xmm9, %zmm8, %zmm2
-; AVX512VL-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,2,8,10,0,2,8,10]
-; AVX512VL-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
-; AVX512VL-NEXT: vpermi2pd %zmm0, %zmm2, %zmm3
-; AVX512VL-NEXT: vinsertf64x4 $0, %ymm1, %zmm3, %zmm0
+; AVX512VL-NEXT: vdivps %xmm4, %xmm0, %xmm0
+; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero
+; AVX512VL-NEXT: vdivps %xmm4, %xmm1, %xmm1
+; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero
+; AVX512VL-NEXT: vdivps %xmm4, %xmm2, %xmm2
+; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero
+; AVX512VL-NEXT: vdivps %xmm4, %xmm3, %xmm3
+; AVX512VL-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2
+; AVX512VL-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0
+; AVX512VL-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,2,8,10,0,2,8,10]
+; AVX512VL-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
+; AVX512VL-NEXT: vpermi2pd %zmm2, %zmm0, %zmm1
+; AVX512VL-NEXT: vmovups (%rdi), %ymm0
+; AVX512VL-NEXT: vdivps (%rsi), %ymm0, %ymm0
+; AVX512VL-NEXT: vinsertf64x4 $0, %ymm0, %zmm1, %zmm0
; AVX512VL-NEXT: vmovupd %zmm0, (%rdx)
; AVX512VL-NEXT: vzeroupper
; AVX512VL-NEXT: retq
More information about the llvm-commits
mailing list