[llvm] 65e86a8 - [X86] combineConcatVectorOps - concat(bitcast(),bitcast()) -> bitcast(concat())
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 22 06:31:34 PDT 2024
Author: Simon Pilgrim
Date: 2024-07-22T14:31:10+01:00
New Revision: 65e86a8f3fb44dc09dc2e08526d69e3a57f63995
URL: https://github.com/llvm/llvm-project/commit/65e86a8f3fb44dc09dc2e08526d69e3a57f63995
DIFF: https://github.com/llvm/llvm-project/commit/65e86a8f3fb44dc09dc2e08526d69e3a57f63995.diff
LOG: [X86] combineConcatVectorOps - concat(bitcast(),bitcast()) -> bitcast(concat())
When concatenating subvector ops together, try to use the pre-bitcasted subop to help expose further combines.
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/pmul.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-7.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
llvm/test/CodeGen/X86/widen_bitcnt.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2959902c78675..f06b962b74df8 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -56165,6 +56165,17 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
SmallVector<SDValue> Subs;
for (SDValue SubOp : SubOps)
Subs.push_back(SubOp.getOperand(I));
+ // Attempt to peek through bitcasts and concat the original subvectors.
+ EVT SubVT = peekThroughBitcasts(Subs[0]).getValueType();
+ if (SubVT.isSimple() && SubVT.isVector()) {
+ EVT ConcatVT =
+ EVT::getVectorVT(*DAG.getContext(), SubVT.getScalarType(),
+ SubVT.getVectorElementCount() * Subs.size());
+ for (SDValue &Sub : Subs)
+ Sub = DAG.getBitcast(SubVT, Sub);
+ return DAG.getBitcast(
+ VT, DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT, Subs));
+ }
return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Subs);
};
auto IsConcatFree = [](MVT VT, ArrayRef<SDValue> SubOps, unsigned Op) {
diff --git a/llvm/test/CodeGen/X86/pmul.ll b/llvm/test/CodeGen/X86/pmul.ll
index fe791e45eff99..2b475644a38cf 100644
--- a/llvm/test/CodeGen/X86/pmul.ll
+++ b/llvm/test/CodeGen/X86/pmul.ll
@@ -964,7 +964,7 @@ define <64 x i8> @mul_v64i8(<64 x i8> %i, <64 x i8> %j) nounwind {
; AVX512F-LABEL: mul_v64i8:
; AVX512F: # %bb.0: # %entry
; AVX512F-NEXT: vextracti64x4 $1, %zmm1, %ymm2
-; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
+; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
; AVX512F-NEXT: vpand %ymm2, %ymm3, %ymm4
; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm5
; AVX512F-NEXT: vpmaddubsw %ymm4, %ymm5, %ymm4
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
index b33cc83ac3f79..c725dcd972cd5 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
@@ -3931,49 +3931,40 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride6_vf32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm9
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm12
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm8
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm12[4],ymm9[4],ymm12[5],ymm9[5],ymm12[6],ymm9[6],ymm12[7],ymm9[7],ymm12[12],ymm9[12],ymm12[13],ymm9[13],ymm12[14],ymm9[14],ymm12[15],ymm9[15]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm4[4],ymm8[5],ymm4[5],ymm8[6],ymm4[6],ymm8[7],ymm4[7],ymm8[12],ymm4[12],ymm8[13],ymm4[13],ymm8[14],ymm4[14],ymm8[15],ymm4[15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm9
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm9[2,1,2,3,6,5,6,7]
; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm10
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm10[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,1,2,3,6,5,6,7]
; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm10[4],ymm8[4],ymm10[5],ymm8[5],ymm10[6],ymm8[6],ymm10[7],ymm8[7],ymm10[12],ymm8[12],ymm10[13],ymm8[13],ymm10[14],ymm8[14],ymm10[15],ymm8[15]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm11[4],ymm9[4],ymm11[5],ymm9[5],ymm11[6],ymm9[6],ymm11[7],ymm9[7],ymm11[12],ymm9[12],ymm11[13],ymm9[13],ymm11[14],ymm9[14],ymm11[15],ymm9[15]
; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm2
; AVX512-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-NEXT: kmovw %eax, %k1
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
; AVX512-NEXT: vextracti64x4 $1, %zmm2, %ymm0
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512-NEXT: vpshufb %ymm4, %ymm1, %ymm3
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm12
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512-NEXT: vpshufb %ymm1, %ymm12, %ymm3
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6],ymm3[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm2[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,1,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm16
; AVX512-NEXT: vmovdqa (%rcx), %ymm2
; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vmovdqa (%rdx), %ymm3
@@ -3983,178 +3974,179 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[1,2,3,3,5,6,7,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
; AVX512-NEXT: vmovdqa (%rsi), %ymm0
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm5[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512-NEXT: vpshufd {{.*#+}} ymm11 = ymm5[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[2],ymm7[2],ymm11[3],ymm7[3],ymm11[8],ymm7[8],ymm11[9],ymm7[9],ymm11[10],ymm7[10],ymm11[11],ymm7[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm11 = ymm5[4],ymm0[4],ymm5[5],ymm0[5],ymm5[6],ymm0[6],ymm5[7],ymm0[7],ymm5[12],ymm0[12],ymm5[13],ymm0[13],ymm5[14],ymm0[14],ymm5[15],ymm0[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm7
-; AVX512-NEXT: vmovdqa32 %zmm6, %zmm7 {%k1}
-; AVX512-NEXT: vextracti64x4 $1, %zmm7, %ymm11
-; AVX512-NEXT: vmovdqa (%r8), %ymm6
-; AVX512-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm11[0],ymm4[1],ymm11[2,3],ymm4[4],ymm11[5,6],ymm4[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm7 = ymm11[0],ymm7[1,2],ymm11[3],ymm7[4,5],ymm11[6],ymm7[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm7[0,1,2,3],zmm4[4,5,6,7]
-; AVX512-NEXT: vmovdqa (%r9), %ymm4
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm21 = ymm7[2,2,2,3]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm22 = ymm7[2,1,2,3]
-; AVX512-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512-NEXT: vmovdqa (%rdx), %xmm14
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm0[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vmovdqa (%rdi), %ymm10
+; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm10[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm10[4],ymm0[4],ymm10[5],ymm0[5],ymm10[6],ymm0[6],ymm10[7],ymm0[7],ymm10[12],ymm0[12],ymm10[13],ymm0[13],ymm10[14],ymm0[14],ymm10[15],ymm0[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[3,3,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}
+; AVX512-NEXT: vextracti64x4 $1, %zmm6, %ymm5
+; AVX512-NEXT: vmovdqa (%r8), %ymm14
+; AVX512-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1],ymm5[2,3],ymm1[4],ymm5[5,6],ymm1[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm6[1,2],ymm5[3],ymm6[4,5],ymm5[6],ymm6[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm17
+; AVX512-NEXT: vmovdqa (%rcx), %xmm5
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm13
+; AVX512-NEXT: vmovdqa (%rdx), %xmm6
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm24 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512-NEXT: vpermt2d %zmm7, %zmm24, %zmm2
-; AVX512-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512-NEXT: vmovdqa (%rdi), %xmm15
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm3[1,1,1,1]
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[2],ymm0[2],ymm5[3],ymm0[3],ymm5[8],ymm0[8],ymm5[9],ymm0[9],ymm5[10],ymm0[10],ymm5[11],ymm0[11]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
+; AVX512-NEXT: vpermt2d %zmm1, %zmm18, %zmm2
+; AVX512-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm10[0],ymm0[0],ymm10[1],ymm0[1],ymm10[2],ymm0[2],ymm10[3],ymm0[3],ymm10[8],ymm0[8],ymm10[9],ymm0[9],ymm10[10],ymm0[10],ymm10[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
; AVX512-NEXT: vextracti64x4 $1, %zmm2, %ymm0
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm5[2],ymm0[3,4],ymm5[5],ymm0[6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm5
-; AVX512-NEXT: vmovdqa (%r8), %xmm0
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-NEXT: vpshufb %xmm6, %xmm0, %xmm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm7[1],ymm2[2,3],ymm7[4],ymm2[5,6],ymm7[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-NEXT: vmovdqa (%r9), %xmm2
-; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm16 = ymm5[0,1,0,1]
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm12[0],ymm9[0],ymm12[1],ymm9[1],ymm12[2],ymm9[2],ymm12[3],ymm9[3],ymm12[8],ymm9[8],ymm12[9],ymm9[9],ymm12[10],ymm9[10],ymm12[11],ymm9[11]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
-; AVX512-NEXT: vpermt2d %zmm7, %zmm24, %zmm9
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm24 = ymm4[2,2,2,2]
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm10[0],ymm8[0],ymm10[1],ymm8[1],ymm10[2],ymm8[2],ymm10[3],ymm8[3],ymm10[8],ymm8[8],ymm10[9],ymm8[9],ymm10[10],ymm8[10],ymm10[11],ymm8[11]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
-; AVX512-NEXT: vmovdqa32 %zmm7, %zmm9 {%k1}
-; AVX512-NEXT: vextracti64x4 $1, %zmm9, %ymm7
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm14[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm7[0,1],ymm1[2],ymm7[3,4],ymm1[5],ymm7[6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm7
-; AVX512-NEXT: vpshufb %xmm6, %xmm7, %xmm6
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1],ymm9[2,3],ymm6[4],ymm9[5,6],ymm6[7]
-; AVX512-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm6[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm5 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
-; AVX512-NEXT: vpermt2d %zmm3, %zmm5, %zmm6
-; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm8 = xmm4[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm3[0,1,0,1]
-; AVX512-NEXT: vmovdqa64 %ymm25, %ymm3
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm9[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
+; AVX512-NEXT: vmovdqa (%r8), %xmm10
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-NEXT: vpshufb %xmm1, %xmm10, %xmm0
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm15
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm8[0],ymm4[0],ymm8[1],ymm4[1],ymm8[2],ymm4[2],ymm8[3],ymm4[3],ymm8[8],ymm4[8],ymm8[9],ymm4[9],ymm8[10],ymm4[10],ymm8[11],ymm4[11]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512-NEXT: vpermt2d %zmm8, %zmm18, %zmm4
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm11[0],ymm9[0],ymm11[1],ymm9[1],ymm11[2],ymm9[2],ymm11[3],ymm9[3],ymm11[8],ymm9[8],ymm11[9],ymm9[9],ymm11[10],ymm9[10],ymm11[11],ymm9[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512-NEXT: vmovdqa32 %zmm8, %zmm4 {%k1}
+; AVX512-NEXT: vextracti64x4 $1, %zmm4, %ymm8
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7]
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm12
+; AVX512-NEXT: vpshufb %xmm1, %xmm12, %xmm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2,3],ymm1[4],ymm4[5,6],ymm1[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm4
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm11
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm1[2,2,2,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm11[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm1[2,1,2,3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm13 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512-NEXT: vpermt2d %zmm0, %zmm18, %zmm1
+; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm14[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = xmm15[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm13[0,1,2,3],zmm0[0,1,0,1]
; AVX512-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-NEXT: kmovw %eax, %k1
-; AVX512-NEXT: vmovdqa32 %zmm6, %zmm4 {%k1}
-; AVX512-NEXT: vextracti64x4 $1, %zmm4, %ymm6
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm7[2,1,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1,2],ymm9[3],ymm6[4,5],ymm9[6],ymm6[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm6[4,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm10[0,0,2,1,4,5,6,7]
-; AVX512-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm7 = xmm10[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm10 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm12 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
-; AVX512-NEXT: vpermt2d %zmm10, %zmm5, %zmm9
-; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm11[0,1,2,1]
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm13 = xmm12[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm13[0],ymm1[1,2],ymm13[3],ymm1[4,5],ymm13[6],ymm1[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
+; AVX512-NEXT: vpbroadcastq %xmm12, %ymm12
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm12
+; AVX512-NEXT: vmovdqa (%r9), %ymm0
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm1[2,2,2,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm0[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm1[2,1,2,3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512-NEXT: vpermt2d %zmm5, %zmm18, %zmm1
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[0,1,2,1]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm5[0,1,0,1]
-; AVX512-NEXT: vmovdqa32 %zmm9, %zmm5 {%k1}
-; AVX512-NEXT: vextracti64x4 $1, %zmm5, %ymm9
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1,2],ymm10[3],ymm9[4,5],ymm10[6],ymm9[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1],ymm0[2],ymm5[3,4],ymm0[5],ymm5[6,7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm9[4,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm6 = xmm7[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm5[0,1,0,1]
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
+; AVX512-NEXT: vextracti64x4 $1, %zmm3, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm10[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
; AVX512-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-NEXT: vmovdqa (%r9), %xmm3
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
+; AVX512-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm11, %ymm11
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm15, %ymm15
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm9
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512-NEXT: vpternlogd $184, %zmm17, %zmm10, %zmm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm11
-; AVX512-NEXT: vpternlogd $184, %zmm20, %zmm10, %zmm11
-; AVX512-NEXT: vinserti64x4 $1, %ymm24, %zmm16, %zmm10
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogd $184, %zmm23, %zmm12, %zmm10
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm3
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm12, %zmm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm1
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-NEXT: vpternlogd $184, %zmm4, %zmm6, %zmm1
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm2
-; AVX512-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm3, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm11, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512-NEXT: vpternlogd $184, %zmm16, %zmm9, %zmm8
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
+; AVX512-NEXT: vpternlogd $184, %zmm17, %zmm9, %zmm13
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm5, %zmm0
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm2
+; AVX512-NEXT: vpternlogd $184, %zmm4, %zmm5, %zmm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm11, %zmm4
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm12, %zmm5, %zmm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm15, %zmm3
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm5, %zmm3
+; AVX512-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm13, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm8, 320(%rax)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -4168,16 +4160,16 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm3
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm2[4],ymm4[5],ymm2[5],ymm4[6],ymm2[6],ymm4[7],ymm2[7],ymm4[12],ymm2[12],ymm4[13],ymm2[13],ymm4[14],ymm2[14],ymm4[15],ymm2[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm24
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm23
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm24
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [2,1,2,3,11,11,11,11]
; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm1
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm12
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm11
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm13[4],ymm12[4],ymm13[5],ymm12[5],ymm13[6],ymm12[6],ymm13[7],ymm12[7],ymm13[12],ymm12[12],ymm13[13],ymm12[13],ymm13[14],ymm12[14],ymm13[15],ymm12[15]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm13[4],ymm11[4],ymm13[5],ymm11[5],ymm13[6],ymm11[6],ymm13[7],ymm11[7],ymm13[12],ymm11[12],ymm13[13],ymm11[13],ymm13[14],ymm11[14],ymm13[15],ymm11[15]
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [5,6,5,6,5,6,7,7]
; AVX512-FCP-NEXT: vpermd %ymm3, %ymm10, %ymm3
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm5 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
@@ -4185,17 +4177,17 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-FCP-NEXT: kmovw %eax, %k1
; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1}
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [8,21,10,11,20,13,14,23]
; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm5
+; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm12, %ymm5
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [8,21,10,11,20,13,14,23]
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm4
-; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm11, %zmm4
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm27
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm17, %ymm1
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[0,1,2,3],zmm4[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm17, %zmm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm16
; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm3
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm3, %ymm1
@@ -4206,13 +4198,13 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [2,2,0,3,10,0,10,11]
; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm2
; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm1
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm1
; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm2
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm9[4],ymm2[5],ymm9[5],ymm2[6],ymm9[6],ymm2[7],ymm9[7],ymm2[12],ymm9[12],ymm2[13],ymm9[13],ymm2[14],ymm9[14],ymm2[15],ymm9[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm28
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm27
; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm0
; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm6
@@ -4225,14 +4217,14 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm10, %ymm2
-; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm11
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm17, %ymm0
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[0,1,2,3],zmm11[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm2
+; AVX512-FCP-NEXT: vpermi2d %ymm2, %ymm0, %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm17, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm12, %zmm17
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm12, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm2
; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm2
; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm14
@@ -4242,120 +4234,120 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm18
; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm6
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm11[0],ymm13[1],ymm11[1],ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[8],ymm11[8],ymm13[9],ymm11[9],ymm13[10],ymm11[10],ymm13[11],ymm11[11]
; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm20
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm13
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm13, %xmm4
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm3, %xmm12
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,2,1,8,9,8,9]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm12
-; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm23 = [1,0,2,2,1,0,2,2]
-; AVX512-FCP-NEXT: # ymm23 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
-; AVX512-FCP-NEXT: vpermd %ymm5, %ymm23, %ymm5
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm5
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,0,2,1,8,9,8,9]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm22, %zmm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
+; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; AVX512-FCP-NEXT: vpermd %ymm8, %ymm4, %ymm8
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
; AVX512-FCP-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm1
; AVX512-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-FCP-NEXT: kmovw %eax, %k2
-; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm12 {%k2}
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [16,9,10,17,12,13,18,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm1
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm8, %zmm5, %zmm1
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT: vpermt2d %ymm8, %ymm22, %ymm12
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm12[0,1,2,3],zmm1[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm8
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm8, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm2
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm12
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm12, %xmm4
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm12[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm1
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512-FCP-NEXT: vpermd %ymm4, %ymm23, %ymm4
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm14 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512-FCP-NEXT: vpbroadcastq %xmm14, %ymm14
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vmovdqa32 %zmm4, %zmm2 {%k2}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm4
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm14 = xmm4[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermi2d %zmm14, %zmm2, %zmm5
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm14 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512-FCP-NEXT: vpermt2d %ymm14, %ymm22, %ymm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm2[0,1,2,3],zmm5[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm11 {%k2}
+; AVX512-FCP-NEXT: vmovdqa %ymm11, %ymm6
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm1
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm11
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm9, %ymm21, %ymm6
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm6, %zmm19
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm11
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm11, %xmm2
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm22, %zmm5
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512-FCP-NEXT: vpermd %ymm2, %ymm4, %ymm2
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm4 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm6 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX512-FCP-NEXT: vpbroadcastq %xmm4, %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm5 {%k2}
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
+; AVX512-FCP-NEXT: vpermi2d %ymm4, %ymm5, %ymm21
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm5
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,0,0,0,8,8,0,9]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm15
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm14 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm4, %zmm14
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm21, %zmm21
; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm5
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm5, %xmm7
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm14 = xmm5[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm14
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm2
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm2[0],ymm9[0],ymm2[1],ymm9[1],ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[8],ymm9[8],ymm2[9],ymm9[9],ymm2[10],ymm9[10],ymm2[11],ymm9[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [1,1,1,1,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm15, %zmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512-FCP-NEXT: vpermd %zmm18, %zmm7, %zmm18
-; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm18 {%k1}
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm5, %xmm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm4, %zmm15
+; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm4
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [1,1,1,1,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm8
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
+; AVX512-FCP-NEXT: vpermd %zmm18, %zmm4, %zmm11
+; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm11 {%k1}
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm11, %ymm22
+; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm18, %ymm22
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [8,9,20,11,12,21,14,15]
; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm10
-; AVX512-FCP-NEXT: vpermt2d %zmm8, %zmm0, %zmm10
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm21, %ymm18
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm5, %xmm8
+; AVX512-FCP-NEXT: vpermt2d %zmm8, %zmm2, %zmm11
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm10 = ymm12[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm5, %xmm8
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm8
-; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm9
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm13[4],xmm3[5],xmm13[5],xmm3[6],xmm13[6],xmm3[7],xmm13[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm15, %zmm3
-; AVX512-FCP-NEXT: vpermd %zmm20, %zmm7, %zmm4
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm4 {%k1}
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm4, %zmm0
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm6, %xmm2
-; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm21, %ymm4
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm12, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm2
+; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm5, %zmm8
+; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm10
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[8],ymm7[8],ymm10[9],ymm7[9],ymm10[10],ymm7[10],ymm10[11],ymm7[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm9, %zmm0
+; AVX512-FCP-NEXT: vpermd %zmm20, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm4 {%k1}
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm0
+; AVX512-FCP-NEXT: vpermi2d %ymm0, %ymm4, %ymm18
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm4
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm1
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm18, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm0
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm18[0,1,2,3],zmm10[0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm8
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 256(%rax)
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm22, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm2, %zmm8
; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm22, %zmm0, %zmm14
-; AVX512-FCP-NEXT: vmovdqa64 %zmm14, (%rax)
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm19, %zmm0, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm21, %zmm0, %zmm15
+; AVX512-FCP-NEXT: vmovdqa64 %zmm15, (%rax)
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm19, %zmm0, %zmm14
+; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT: vpternlogd $184, %zmm17, %zmm0, %zmm1
@@ -4370,261 +4362,253 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-LABEL: store_i16_stride6_vf32:
; AVX512DQ: # %bb.0:
; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm2
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm9
; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm3
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm25
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm27
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm26
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm2 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm2, %zmm0
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm8
; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm4
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm10
; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm5
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm28
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm27
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm28
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
; AVX512DQ-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-NEXT: kmovw %eax, %k1
; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm11
; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm4
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1,2],ymm3[3],ymm0[4,5],ymm3[6],ymm0[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm31
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm29
; AVX512DQ-NEXT: vpbroadcastq %xmm3, %ymm3
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm4
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm0, %ymm26
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm30
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm19 = ymm0[0,0,2,1]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm16
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm25
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm24
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm2, %zmm0
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm12[0,1,2,1]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm8[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[0,1,2,1]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm10[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm31
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[0,1,0,1]
; AVX512DQ-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm11[2,1,3,3,4,5,6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm17
+; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero
; AVX512DQ-NEXT: vpbroadcastq %xmm2, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm4[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm0, %ymm22
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm4[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm16
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm0[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm17
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm13[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm7
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm14
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm13
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm2[2,2,2,2]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm7[4],ymm15[4],ymm7[5],ymm15[5],ymm7[6],ymm15[6],ymm7[7],ymm15[7],ymm7[12],ymm15[12],ymm7[13],ymm15[13],ymm7[14],ymm15[14],ymm7[15],ymm15[15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[1,2,3,3,5,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-NEXT: kmovw %eax, %k1
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm6
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm6, %ymm1
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm6, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm22
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,1,2,3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm23 = ymm0[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm9 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm9[0],ymm2[0],ymm9[1],ymm2[1],ymm9[2],ymm2[2],ymm9[3],ymm2[3],ymm9[8],ymm2[8],ymm9[9],ymm2[9],ymm9[10],ymm2[10],ymm9[11],ymm2[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm1[4],ymm3[4],ymm1[5],ymm3[5],ymm1[6],ymm3[6],ymm1[7],ymm3[7],ymm1[12],ymm3[12],ymm1[13],ymm3[13],ymm1[14],ymm3[14],ymm1[15],ymm3[15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm0
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm9, %zmm2, %zmm0 {%k1}
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm2, %ymm9
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm10
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6],ymm9[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm10[0],ymm0[1,2],ymm10[3],ymm0[4,5],ymm10[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm0[0,1,2,3],zmm9[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm9
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm9[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,1,2,3]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm0[3,3,3,3]
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm8 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[1,2,3,3,5,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm18, %zmm12
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm12 {%k1}
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm1, %ymm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm12, %ymm8
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3],ymm5[4],ymm8[5,6],ymm5[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0],ymm12[1,2],ymm8[3],ymm12[4,5],ymm8[6],ymm12[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm18
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm12
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm5, %ymm19
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm20 = ymm5[0,0,2,1]
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm5
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm8, %ymm21
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm5[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm23 = ymm8[0,0,2,1]
+; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm8
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm24 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
+; AVX512DQ-NEXT: vpermt2d %zmm8, %zmm24, %zmm0
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm2
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm11, %xmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm2
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,2,2,3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm0[4],xmm11[5],xmm0[5],xmm11[6],xmm0[6],xmm11[7],xmm0[7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm9[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm25 = ymm10[2,1,2,3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm10 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm10, %zmm3
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm12[4],xmm15[5],xmm12[5],xmm15[6],xmm12[6],xmm15[7],xmm12[7]
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm11 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm5[2,1,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm4, %zmm0, %zmm3 {%k1}
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm3, %ymm2
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm17, %xmm4
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3],ymm4[4],ymm3[5,6],ymm4[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm16[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm9
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm7[0],ymm15[0],ymm7[1],ymm15[1],ymm7[2],ymm15[2],ymm7[3],ymm15[3],ymm7[8],ymm15[8],ymm7[9],ymm15[9],ymm7[10],ymm15[10],ymm7[11],ymm15[11]
+; AVX512DQ-NEXT: vpermt2d %zmm8, %zmm24, %zmm7
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm9
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm9, %zmm8, %zmm7 {%k1}
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm7, %ymm8
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1],ymm6[2],ymm8[3,4],ymm6[5],ymm8[6,7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm8
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm8, %xmm3
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm5
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm4 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm4 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11]
-; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm10, %zmm5
-; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm10
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm10[4],xmm4[5],xmm10[5],xmm4[6],xmm10[6],xmm4[7],xmm10[7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm3 = ymm7[0],ymm3[1],ymm7[2,3],ymm3[4],ymm7[5,6],ymm3[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm6
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm7 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm7, %zmm4, %zmm5 {%k1}
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm5, %ymm4
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm6
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm6, %xmm0
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm4 = xmm30[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm9 = xmm12[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm19, %zmm26, %zmm6
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm18, %zmm22, %zmm8
-; AVX512DQ-NEXT: vpternlogd $184, %zmm21, %zmm7, %zmm8
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm7
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm20, %zmm9, %zmm7
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm10
-; AVX512DQ-NEXT: vpternlogd $184, %zmm23, %zmm9, %zmm10
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm3, %zmm1
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm2
-; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm2
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm10, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm7, 320(%rax)
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm10
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm16, %zmm11, %zmm10
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm23, %zmm21, %zmm12
+; AVX512DQ-NEXT: vpternlogd $184, %zmm17, %zmm11, %zmm12
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm22, %zmm4, %zmm0
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512DQ-NEXT: vpternlogd $184, %zmm18, %zmm4, %zmm7
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm4
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm5, %zmm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
+; AVX512DQ-NEXT: vpternlogd $184, %zmm3, %zmm5, %zmm1
+; AVX512DQ-NEXT: vmovdqa64 %zmm1, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm4, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm12, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm10, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, 320(%rax)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i16_stride6_vf32:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm14[0],ymm1[1],ymm14[1],ymm1[2],ymm14[2],ymm1[3],ymm14[3],ymm1[8],ymm14[8],ymm1[9],ymm14[9],ymm1[10],ymm14[10],ymm1[11],ymm14[11]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm31
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm25
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1
; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [1,1,1,1,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm27
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm3
; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
@@ -4633,63 +4617,62 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm21 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm21, %zmm19
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
+; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm20, %zmm19
; AVX512DQ-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FCP-NEXT: kmovw %eax, %k1
; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm1, %zmm19 {%k1}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm23
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm23
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm0, %xmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm23
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm17, %ymm23
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm21 = [8,9,20,11,12,21,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm9
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm4
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm30
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm29
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm28
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm6, %ymm19
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm7
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,0,1,0,10,10,0]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm12
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm12
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm7
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm21, %zmm19
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm8
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm28
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm10
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm20, %zmm0
-; AVX512DQ-FCP-NEXT: vpermd %zmm3, %zmm21, %zmm20
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm18, %zmm13
+; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm20, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm13, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm12, %xmm3
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm3, %ymm20, %ymm17
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm20, %zmm17
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm11, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm6, %ymm20
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm31
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm5 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm21, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm24
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm5
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm13
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm13
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm15
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm15[0],ymm0[0],ymm15[1],ymm0[1],ymm15[2],ymm0[2],ymm15[3],ymm0[3],ymm15[8],ymm0[8],ymm15[9],ymm0[9],ymm15[10],ymm0[10],ymm15[11],ymm0[11]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm14
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm14[0],ymm0[0],ymm14[1],ymm0[1],ymm14[2],ymm0[2],ymm14[3],ymm0[3],ymm14[8],ymm0[8],ymm14[9],ymm0[9],ymm14[10],ymm0[10],ymm14[11],ymm0[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,1,2,3,11,11,11,11]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
@@ -4701,126 +4684,126 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm1, %zmm4, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm9
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm9
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm4, %ymm9
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm3
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm15, %zmm4, %zmm9
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm22, %ymm0
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm0[0,1,2,3],zmm9[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm14, %ymm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm22, %zmm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm18
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm9, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm9, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm3
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm9[4],ymm14[4],ymm9[5],ymm14[5],ymm9[6],ymm14[6],ymm9[7],ymm14[7],ymm9[12],ymm14[12],ymm9[13],ymm14[13],ymm9[14],ymm14[14],ymm9[15],ymm14[15]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm9[4],ymm1[5],ymm9[5],ymm1[6],ymm9[6],ymm1[7],ymm9[7],ymm1[12],ymm9[12],ymm1[13],ymm9[13],ymm1[14],ymm9[14],ymm1[15],ymm9[15]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm6
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm3
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm14
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,0,3,10,0,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm14
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm14
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm1
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm1[4],ymm14[4],ymm1[5],ymm14[5],ymm1[6],ymm14[6],ymm1[7],ymm14[7],ymm1[12],ymm14[12],ymm1[13],ymm14[13],ymm1[14],ymm14[14],ymm1[15],ymm14[15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm6, %ymm21, %ymm6
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm15 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm14 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm14 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[8],ymm15[8],ymm14[9],ymm15[9],ymm14[10],ymm15[10],ymm14[11],ymm15[11]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm6, %zmm14, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm9
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm9[4],ymm1[5],ymm9[5],ymm1[6],ymm9[6],ymm1[7],ymm9[7],ymm1[12],ymm9[12],ymm1[13],ymm9[13],ymm1[14],ymm9[14],ymm1[15],ymm9[15]
+; AVX512DQ-FCP-NEXT: vpermd %ymm3, %ymm21, %ymm3
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm9[0],ymm1[1],ymm9[1],ymm1[2],ymm9[2],ymm1[3],ymm9[3],ymm1[8],ymm9[8],ymm1[9],ymm9[9],ymm1[10],ymm9[10],ymm1[11],ymm9[11]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm3, %zmm1, %zmm0 {%k1}
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm22, %ymm0
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm0[0,1,2,3],zmm4[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm10, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm1
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm22, %zmm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm8, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm5, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm11, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm4
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,2,1,8,9,8,9]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm3
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm9
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm8
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX512DQ-FCP-NEXT: vpermd %ymm6, %ymm4, %ymm6
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm7, %ymm7
; AVX512DQ-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FCP-NEXT: kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm8, %zmm6, %zmm3 {%k1}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm6 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm11[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm9
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm8, %zmm6, %zmm9
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm7, %zmm6, %zmm3 {%k1}
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, %ymm7
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm12[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm9, %zmm8, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm9, %ymm7
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm10[0],xmm7[1],xmm10[1],xmm7[2],xmm10[2],xmm7[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm5, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm5
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [0,0,0,0,8,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm7
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm10[4],xmm7[5],xmm10[5],xmm7[6],xmm10[6],xmm7[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm11[0],xmm14[1],xmm11[1],xmm14[2],xmm11[2],xmm14[3],xmm11[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm5, %zmm10
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm5 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm8, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm9[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm7 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm8 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm7, %zmm4, %zmm10 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm7
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm7, %xmm4
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm11
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm10, %zmm6
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm28[0],zero,xmm28[1],zero,xmm28[2],zero,xmm28[3],zero
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm5, %ymm10
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm10[0,1,2,3],zmm6[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX512DQ-FCP-NEXT: vpermd %ymm5, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm11 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm5[0],xmm11[1],xmm5[1],xmm11[2],xmm5[2],xmm11[3],xmm5[3]
+; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm5, %ymm5
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm5, %zmm4, %zmm6 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm4, %ymm6, %ymm9
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm4, %zmm8, %zmm6
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm6
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm6, %xmm5
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm9, %zmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm6
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm4, %zmm5, %zmm6
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm5, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm1, %zmm3, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm18, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm20[0,1,2,3],zmm17[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm5, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm2, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 128(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm18, %zmm2, %zmm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm20, %zmm17, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm1, %zmm13
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm19[0,1,2,3],zmm23[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm1, %zmm12
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 64(%rax)
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm23, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm1, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -8334,901 +8317,875 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride6_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $600, %rsp # imm = 0x258
+; AVX512-NEXT: subq $392, %rsp # imm = 0x188
; AVX512-NEXT: vmovdqa 96(%rcx), %ymm9
; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa 96(%rdx), %ymm6
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 96(%rdx), %ymm4
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm6[4],ymm9[4],ymm6[5],ymm9[5],ymm6[6],ymm9[6],ymm6[7],ymm9[7],ymm6[12],ymm9[12],ymm6[13],ymm9[13],ymm6[14],ymm9[14],ymm6[15],ymm9[15]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm9[4],ymm4[5],ymm9[5],ymm4[6],ymm9[6],ymm4[7],ymm9[7],ymm4[12],ymm9[12],ymm4[13],ymm9[13],ymm4[14],ymm9[14],ymm4[15],ymm9[15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-NEXT: vmovdqa 96(%rsi), %ymm10
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[2,1,2,3,6,5,6,7]
; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa 96(%rdi), %ymm8
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-NEXT: vmovdqa 96(%rdi), %ymm6
+; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm8[4],ymm10[4],ymm8[5],ymm10[5],ymm8[6],ymm10[6],ymm8[7],ymm10[7],ymm8[12],ymm10[12],ymm8[13],ymm10[13],ymm8[14],ymm10[14],ymm8[15],ymm10[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm6[4],ymm10[4],ymm6[5],ymm10[5],ymm6[6],ymm10[6],ymm6[7],ymm10[7],ymm6[12],ymm10[12],ymm6[13],ymm10[13],ymm6[14],ymm10[14],ymm6[15],ymm10[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-NEXT: kmovw %eax, %k1
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512-NEXT: vmovdqa 96(%r8), %ymm5
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512-NEXT: vpshufb %ymm7, %ymm5, %ymm2
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 96(%r9), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-NEXT: vmovdqa 96(%r8), %ymm3
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512-NEXT: vpshufb %ymm14, %ymm3, %ymm5
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3],ymm5[4],ymm0[5,6],ymm5[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 64(%rcx), %ymm2
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa 64(%rdx), %ymm3
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 64(%rcx), %ymm5
+; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 64(%rdx), %ymm7
+; AVX512-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm21
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm5[4],ymm7[5],ymm5[5],ymm7[6],ymm5[6],ymm7[7],ymm5[7],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-NEXT: vmovdqa 64(%rsi), %ymm3
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vmovdqa 64(%rsi), %ymm7
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,1,2,3,6,5,6,7]
; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa 64(%rdi), %ymm4
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-NEXT: vmovdqa 64(%rdi), %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm1[0],ymm5[1],ymm1[1],ymm5[2],ymm1[2],ymm5[3],ymm1[3],ymm5[8],ymm1[8],ymm5[9],ymm1[9],ymm5[10],ymm1[10],ymm5[11],ymm1[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm16
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm2[4],ymm7[4],ymm2[5],ymm7[5],ymm2[6],ymm7[6],ymm2[7],ymm7[7],ymm2[12],ymm7[12],ymm2[13],ymm7[13],ymm2[14],ymm7[14],ymm2[15],ymm7[15]
+; AVX512-NEXT: vmovdqa64 %ymm7, %ymm20
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512-NEXT: vmovdqa 64(%r8), %ymm11
-; AVX512-NEXT: vpshufb %ymm7, %ymm11, %ymm2
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 64(%r9), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-NEXT: vmovdqa 64(%r8), %ymm12
+; AVX512-NEXT: vpshufb %ymm14, %ymm12, %ymm5
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3],ymm5[4],ymm0[5,6],ymm5[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm13
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm12
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm5
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm5[4],ymm7[5],ymm5[5],ymm7[6],ymm5[6],ymm7[7],ymm5[7],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15]
+; AVX512-NEXT: vmovdqa64 %ymm7, %ymm29
+; AVX512-NEXT: vmovdqa64 %ymm5, %ymm18
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm4
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm14
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm14[4],ymm4[4],ymm14[5],ymm4[5],ymm14[6],ymm4[6],ymm14[7],ymm4[7],ymm14[12],ymm4[12],ymm14[13],ymm4[13],ymm14[14],ymm4[14],ymm14[15],ymm4[15]
-; AVX512-NEXT: vmovdqa64 %ymm14, %ymm23
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm24
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm7
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vmovdqa %ymm7, %ymm8
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm2[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm7[0],ymm5[0],ymm7[1],ymm5[1],ymm7[2],ymm5[2],ymm7[3],ymm5[3],ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm5[2,1,2,3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm2[4],ymm8[4],ymm2[5],ymm8[5],ymm2[6],ymm8[6],ymm2[7],ymm8[7],ymm2[12],ymm8[12],ymm2[13],ymm8[13],ymm2[14],ymm8[14],ymm2[15],ymm8[15]
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512-NEXT: vmovdqa64 %ymm8, %ymm16
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
; AVX512-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512-NEXT: vpshufb %ymm7, %ymm3, %ymm0
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm15
+; AVX512-NEXT: vpshufb %ymm14, %ymm15, %ymm0
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm2
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm5
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6],ymm0[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,1,2,3]
-; AVX512-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512-NEXT: vpsrldq {{.*#+}} ymm3 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm25 = ymm0[2,2,2,2]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm20 = ymm3[1,2,3,3,5,6,7,7]
-; AVX512-NEXT: vmovdqa (%rsi), %ymm4
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm15[0],ymm0[0],ymm15[1],ymm0[1],ymm15[2],ymm0[2],ymm15[3],ymm0[3],ymm15[8],ymm0[8],ymm15[9],ymm0[9],ymm15[10],ymm0[10],ymm15[11],ymm0[11]
+; AVX512-NEXT: vmovdqa (%rcx), %ymm8
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm0 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vmovdqa (%rdx), %ymm13
+; AVX512-NEXT: vpsrldq {{.*#+}} ymm1 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,2,2,2]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm13[4],ymm8[4],ymm13[5],ymm8[5],ymm13[6],ymm8[6],ymm13[7],ymm8[7],ymm13[12],ymm8[12],ymm13[13],ymm8[13],ymm13[14],ymm8[14],ymm13[15],ymm8[15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[1,2,3,3,5,6,7,7]
+; AVX512-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512-NEXT: vpshufd {{.*#+}} ymm11 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[8],ymm0[8],ymm11[9],ymm0[9],ymm11[10],ymm0[10],ymm11[11],ymm0[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm15 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm20, %zmm25, %zmm20
-; AVX512-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm15
-; AVX512-NEXT: vmovdqa32 %zmm20, %zmm15 {%k1}
-; AVX512-NEXT: vmovdqa (%r8), %ymm0
-; AVX512-NEXT: vpshufb %ymm7, %ymm0, %ymm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
-; AVX512-NEXT: vextracti64x4 $1, %zmm15, %ymm14
-; AVX512-NEXT: vpblendd {{.*#+}} ymm7 = ymm14[0],ymm7[1],ymm14[2,3],ymm7[4],ymm14[5,6],ymm7[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm11 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm19, %zmm7
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
+; AVX512-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}
+; AVX512-NEXT: vmovdqa (%r8), %ymm7
+; AVX512-NEXT: vpshufb %ymm14, %ymm7, %ymm11
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
+; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm14
+; AVX512-NEXT: vpblendd {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3],ymm11[4],ymm14[5,6],ymm11[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm7[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0],ymm15[1,2],ymm14[3],ymm15[4,5],ymm14[6],ymm15[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm14[0,1,2,3],zmm7[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm29 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm14[0],ymm0[1,2],ymm14[3],ymm0[4,5],ymm14[6],ymm0[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm8[0],ymm13[1],ymm8[1],ymm13[2],ymm8[2],ymm13[3],ymm8[3],ymm13[8],ymm8[8],ymm13[9],ymm8[9],ymm13[10],ymm8[10],ymm13[11],ymm8[11]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm23 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
+; AVX512-NEXT: vmovdqa (%rcx), %xmm13
; AVX512-NEXT: vmovdqa (%rdx), %xmm14
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm14[4],xmm7[4],xmm14[5],xmm7[5],xmm14[6],xmm7[6],xmm14[7],xmm7[7]
-; AVX512-NEXT: vmovdqa64 %xmm14, %xmm26
-; AVX512-NEXT: vpermt2d %zmm2, %zmm29, %zmm1
-; AVX512-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm14[4],xmm2[5],xmm14[5],xmm2[6],xmm14[6],xmm2[7],xmm14[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512-NEXT: vmovdqa32 %zmm2, %zmm1 {%k1}
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm2
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vmovdqa (%r8), %xmm2
-; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-NEXT: vmovdqa %xmm3, %xmm14
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512-NEXT: vpermt2d %zmm8, %zmm23, %zmm0
+; AVX512-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[8],ymm2[8],ymm5[9],ymm2[9],ymm5[10],ymm2[10],ymm5[11],ymm2[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm1
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm5
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1],ymm1[2],ymm5[3,4],ymm1[5],ymm5[6,7]
+; AVX512-NEXT: vmovdqa (%r8), %xmm8
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm5
+; AVX512-NEXT: vmovdqa %xmm2, %xmm11
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3],ymm5[4],ymm0[5,6],ymm5[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[2],ymm9[2],ymm6[3],ymm9[3],ymm6[8],ymm9[8],ymm6[9],ymm9[9],ymm6[10],ymm9[10],ymm6[11],ymm9[11]
-; AVX512-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512-NEXT: vmovdqa 96(%rdx), %xmm4
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-NEXT: vpermt2d %zmm0, %zmm29, %zmm6
-; AVX512-NEXT: vmovdqa 96(%rsi), %xmm1
-; AVX512-NEXT: vmovdqa 96(%rdi), %xmm0
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[8],ymm10[8],ymm8[9],ymm10[9],ymm8[10],ymm10[10],ymm8[11],ymm10[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
-; AVX512-NEXT: vmovdqa32 %zmm2, %zmm6 {%k1}
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm5[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512-NEXT: vextracti64x4 $1, %zmm6, %ymm5
-; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm5
-; AVX512-NEXT: vmovdqa 96(%r8), %xmm2
-; AVX512-NEXT: vpshufb %xmm14, %xmm2, %xmm8
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2,3],ymm8[4],ymm6[5,6],ymm8[7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm6[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm4[0],ymm9[0],ymm4[1],ymm9[1],ymm4[2],ymm9[2],ymm4[3],ymm9[3],ymm4[8],ymm9[8],ymm4[9],ymm9[9],ymm4[10],ymm9[10],ymm4[11],ymm9[11]
+; AVX512-NEXT: vmovdqa 96(%rcx), %xmm4
+; AVX512-NEXT: vmovdqa 96(%rdx), %xmm5
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vpermt2d %zmm0, %zmm23, %zmm7
+; AVX512-NEXT: vmovdqa 96(%rsi), %xmm0
+; AVX512-NEXT: vmovdqa 96(%rdi), %xmm2
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm10[0],ymm6[1],ymm10[1],ymm6[2],ymm10[2],ymm6[3],ymm10[3],ymm6[8],ymm10[8],ymm6[9],ymm10[9],ymm6[10],ymm10[10],ymm6[11],ymm10[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm7 {%k1}
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512-NEXT: vextracti64x4 $1, %zmm7, %ymm3
+; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7]
+; AVX512-NEXT: vmovdqa 96(%r8), %xmm1
+; AVX512-NEXT: vpshufb %xmm11, %xmm1, %xmm6
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5,6],ymm6[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm3
+; AVX512-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT: vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm31 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
-; AVX512-NEXT: vpermt2d %zmm3, %zmm31, %zmm5
-; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpermt2d %zmm4, %zmm31, %zmm3
; AVX512-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[0,1,2,1]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,0,1]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[0,1,0,1]
; AVX512-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-NEXT: kmovw %eax, %k2
-; AVX512-NEXT: vmovdqa32 %zmm5, %zmm0 {%k2}
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[2,1,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512-NEXT: vmovdqa32 %zmm3, %zmm0 {%k2}
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm3
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3],ymm3[4,5],ymm1[6],ymm3[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-NEXT: vmovdqa 64(%rcx), %xmm2
-; AVX512-NEXT: vmovdqa 64(%rdx), %xmm3
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
+; AVX512-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm22
+; AVX512-NEXT: vmovdqa 64(%rcx), %xmm3
+; AVX512-NEXT: vmovdqa 64(%rdx), %xmm2
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512-NEXT: vpermt2d %zmm0, %zmm31, %zmm5
-; AVX512-NEXT: vmovdqa 64(%rsi), %xmm1
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vmovdqa 64(%rdi), %xmm4
-; AVX512-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm0[0,1,0,1]
-; AVX512-NEXT: vmovdqa32 %zmm5, %zmm6 {%k2}
-; AVX512-NEXT: vmovdqa 64(%r8), %xmm0
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512-NEXT: vpermt2d %zmm0, %zmm31, %zmm1
+; AVX512-NEXT: vmovdqa 64(%rsi), %xmm6
+; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm6[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vmovdqa 64(%rdi), %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512-NEXT: vextracti64x4 $1, %zmm6, %ymm8
-; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm8[1,2],ymm5[3],ymm8[4,5],ymm5[6],ymm8[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1],ymm8[2],ymm6[3,4],ymm8[5],ymm6[6,7]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm6[0,1,2,3],zmm5[4,5,6,7]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm3
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm5
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[2],ymm5[2],ymm3[3],ymm5[3],ymm3[8],ymm5[8],ymm3[9],ymm5[9],ymm3[10],ymm5[10],ymm3[11],ymm5[11]
-; AVX512-NEXT: vpermt2d %zmm2, %zmm29, %zmm3
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
-; AVX512-NEXT: vmovdqa64 %ymm16, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm4
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm5[0,1,2,3],zmm4[0,1,0,1]
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm4 {%k2}
+; AVX512-NEXT: vmovdqa 64(%r8), %xmm1
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512-NEXT: vextracti64x4 $1, %zmm4, %ymm7
+; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm7[1,2],ymm5[3],ymm7[4,5],ymm5[6],ymm7[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
+; AVX512-NEXT: vpbroadcastq %xmm7, %ymm7
+; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm28
+; AVX512-NEXT: vmovdqa 96(%r9), %ymm7
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm24 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm25 = ymm5[2,1,2,3]
+; AVX512-NEXT: vmovdqa 64(%r9), %ymm4
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm26 = ymm5[2,2,2,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm4[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm21
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm27 = ymm5[2,1,2,3]
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm9[2,2,2,3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[8],mem[8],ymm3[9],mem[9],ymm3[10],mem[10],ymm3[11],mem[11]
+; AVX512-NEXT: vpermt2d %zmm2, %zmm23, %zmm3
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
+; AVX512-NEXT: vmovdqa64 %ymm20, %ymm6
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm6[0],ymm2[1],ymm6[1],ymm2[2],ymm6[2],ymm2[3],ymm6[3],ymm2[8],ymm6[8],ymm2[9],ymm6[9],ymm2[10],ymm6[10],ymm2[11],ymm6[11]
; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
; AVX512-NEXT: vextracti64x4 $1, %zmm3, %ymm2
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
-; AVX512-NEXT: vmovdqa %xmm14, %xmm4
-; AVX512-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6],ymm0[7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-NEXT: vmovdqa (%r9), %ymm0
-; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512-NEXT: vpermq {{.*#+}} ymm20 = ymm1[2,2,2,3]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm0[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm16 = ymm1[2,1,2,3]
-; AVX512-NEXT: vmovdqa (%r9), %xmm11
-; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm11[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,1,0,1]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm0[2,2,2,2]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
+; AVX512-NEXT: vmovdqa %xmm11, %xmm4
+; AVX512-NEXT: vpshufb %xmm11, %xmm1, %xmm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm20
; AVX512-NEXT: vmovdqa 32(%rcx), %xmm0
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm13
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-NEXT: vpermt2d %zmm1, %zmm29, %zmm6
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm24, %ymm3
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm1
+; AVX512-NEXT: vmovdqa64 %ymm29, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm3
; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512-NEXT: vmovdqa32 %zmm8, %zmm6 {%k1}
-; AVX512-NEXT: vmovdqa64 %ymm17, %ymm2
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
-; AVX512-NEXT: vextracti64x4 $1, %zmm6, %ymm9
-; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7]
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm8
-; AVX512-NEXT: vpshufb %xmm4, %xmm8, %xmm9
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm6[0],ymm9[1],ymm6[2,3],ymm9[4],ymm6[5,6],ymm9[7]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 32-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm29 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm6 # 64-byte Folded Reload
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm23 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm23 # 64-byte Folded Reload
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm22, %zmm2, %zmm12
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm14
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm9[0,1,2,3],zmm14[4,5,6,7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-NEXT: vpermt2d %zmm2, %zmm23, %zmm9
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[1,1,1,1]
+; AVX512-NEXT: vmovdqa64 %ymm16, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm17, %ymm11
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm11[0],ymm6[0],ymm11[1],ymm6[1],ymm11[2],ymm6[2],ymm11[3],ymm6[3],ymm11[8],ymm6[8],ymm11[9],ymm6[9],ymm11[10],ymm6[10],ymm11[11],ymm6[11]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm11
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm5[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm17 = ymm10[2,1,2,3]
+; AVX512-NEXT: vmovdqa32 %zmm11, %zmm9 {%k1}
+; AVX512-NEXT: vextracti64x4 $1, %zmm9, %ymm11
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm12[2],ymm11[3,4],ymm12[5],ymm11[6,7]
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm15
+; AVX512-NEXT: vpshufb %xmm4, %xmm15, %xmm12
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0],ymm12[1],ymm9[2,3],ymm12[4],ymm9[5,6],ymm12[7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm19
+; AVX512-NEXT: vmovdqa (%r9), %ymm11
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm13 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
-; AVX512-NEXT: vpermt2d %zmm0, %zmm31, %zmm9
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm15[0,1,2,1]
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512-NEXT: vpermt2d %zmm0, %zmm31, %zmm12
+; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm2[0,1,2,1]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = xmm1[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-NEXT: vmovdqa 96(%r9), %xmm13
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm24 = ymm14[0,1,0,1]
-; AVX512-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm14 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm17 = ymm14[2,2,2,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[0,1,0,1]
-; AVX512-NEXT: vmovdqa32 %zmm9, %zmm0 {%k2}
-; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[2,1,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm9[0],ymm1[1,2],ymm9[3],ymm1[4,5],ymm9[6],ymm1[7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[0,0,2,1,4,5,6,7]
-; AVX512-NEXT: vpbroadcastq %xmm9, %ymm18
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm19 = ymm9[0,0,2,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm8[2],ymm0[3,4],ymm8[5],ymm0[6,7]
-; AVX512-NEXT: vmovdqa 64(%r9), %xmm8
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-NEXT: vmovdqa64 %xmm26, %xmm2
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm1 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpsrldq {{.*#+}} xmm7 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm7 = xmm8[0,0,2,1,4,5,6,7]
-; AVX512-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512-NEXT: vpermt2d %zmm1, %zmm31, %zmm0
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[0,1,2,1]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm11[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm16 = ymm2[2,1,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
+; AVX512-NEXT: vmovdqa32 %zmm12, %zmm1 {%k2}
+; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm3
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm15[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm12[0],ymm3[1,2],ymm12[3],ymm3[4,5],ymm12[6],ymm3[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm12 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero
+; AVX512-NEXT: vpbroadcastq %xmm12, %ymm12
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm12[2],ymm1[3,4],ymm12[5],ymm1[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm15
+; AVX512-NEXT: vmovdqa (%r9), %xmm3
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm23 = ymm1[0,1,0,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm1[2,2,2,2]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm11 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpsrldq {{.*#+}} xmm12 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX512-NEXT: vpermt2d %zmm11, %zmm31, %zmm1
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,7,6,5]
; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX512-NEXT: vpshufd {{.*#+}} xmm9 = xmm2[0,1,2,1]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,5]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm13 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm14[0,1,2,3],zmm1[0,1,0,1]
-; AVX512-NEXT: vmovdqa32 %zmm0, %zmm1 {%k2}
-; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm14 = xmm3[2,1,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm14[0],ymm0[1,2],ymm14[3],ymm0[4,5],ymm14[6],ymm0[7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm14 = xmm15[2,3,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,2,1,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
-; AVX512-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm2 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm26 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
-; AVX512-NEXT: vpbroadcastq %xmm26, %ymm3
-; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm15[0,0,2,1,4,5,6,7]
-; AVX512-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm11[0,0,2,1,4,5,6,7]
-; AVX512-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm12 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm4
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm4 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm5 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm24, %zmm16
-; AVX512-NEXT: vpternlogd $184, %zmm27, %zmm10, %zmm16
-; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
-; AVX512-NEXT: vpternlogd $184, %zmm25, %zmm10, %zmm8
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm2
-; AVX512-NEXT: vpternlogd $184, %zmm22, %zmm10, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm10
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-NEXT: vpternlogd $184, %zmm28, %zmm13, %zmm10
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512-NEXT: vpternlogd $184, %zmm30, %zmm13, %zmm7
-; AVX512-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
-; AVX512-NEXT: vpternlogd $184, %zmm21, %zmm13, %zmm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm1, %zmm1
-; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm13, %zmm1
+; AVX512-NEXT: vpshufd {{.*#+}} xmm12 = xmm2[0,1,2,1]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,7,6,5]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm12[0,1,2,3],zmm11[0,1,0,1]
+; AVX512-NEXT: vmovdqa32 %zmm1, %zmm11 {%k2}
+; AVX512-NEXT: vextracti64x4 $1, %zmm11, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm12[0],ymm1[1,2],ymm12[3],ymm1[4,5],ymm12[6],ymm1[7]
+; AVX512-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
+; AVX512-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm11[0,1],ymm8[2],ymm11[3,4],ymm8[5],ymm11[6,7]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm1
+; AVX512-NEXT: vmovdqa 96(%r9), %xmm8
+; AVX512-NEXT: vpshufd {{.*#+}} xmm11 = xmm8[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm4 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm8[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm12, %ymm12
+; AVX512-NEXT: vmovdqa 64(%r9), %xmm13
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm14 = xmm13[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm14, %ymm14
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm13[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm24
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm24 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm27, %zmm26, %zmm26
+; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
+; AVX512-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm2
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm7 = ymm2[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm10 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[0,2,2,1,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
+; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm26 # 64-byte Folded Reload
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm9, %ymm9
+; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm30, %zmm17
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7]
+; AVX512-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm17 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm16, %zmm18, %zmm16
+; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm16 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm29, %zmm23, %zmm18
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm23, %zmm18 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm11, %zmm4
+; AVX512-NEXT: vpternlogd $184, (%rsp), %zmm23, %zmm4 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7
+; AVX512-NEXT: vpternlogd $184, %zmm20, %zmm23, %zmm7
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm10, %zmm5
+; AVX512-NEXT: vpternlogd $184, %zmm19, %zmm23, %zmm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm12, %zmm8
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm22, %zmm10, %zmm8
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm14, %zmm6
+; AVX512-NEXT: vpternlogd $184, %zmm28, %zmm10, %zmm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
+; AVX512-NEXT: vpternlogd $184, %zmm15, %zmm10, %zmm0
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm10, %zmm2
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm8, 448(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm10, 576(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm16, 640(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm5, 64(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm12, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm23, 512(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm6, 704(%rax)
-; AVX512-NEXT: addq $600, %rsp # imm = 0x258
+; AVX512-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm7, 448(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm6, 384(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm8, 576(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm4, 640(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm18, 64(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm17, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm26, 512(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm24, 704(%rax)
+; AVX512-NEXT: addq $392, %rsp # imm = 0x188
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i16_stride6_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $1256, %rsp # imm = 0x4E8
-; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm3
-; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %ymm4
-; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm2
+; AVX512-FCP-NEXT: subq $1240, %rsp # imm = 0x4D8
+; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm1
+; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %ymm2
+; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm6
; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm5
; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %xmm9
-; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm11
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm12
+; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm10
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm13
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512-FCP-NEXT: vmovdqa %xmm8, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm12
; AVX512-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm6
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512-FCP-NEXT: vmovdqa %xmm7, (%rsp) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm13
-; AVX512-FCP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm10
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[8],ymm2[8],ymm5[9],ymm2[9],ymm5[10],ymm2[10],ymm5[11],ymm2[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm10, %xmm19
-; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm23
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm10
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[2],ymm7[2],ymm10[3],ymm7[3],ymm10[8],ymm7[8],ymm10[9],ymm7[9],ymm10[10],ymm7[10],ymm10[11],ymm7[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm11
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[8],ymm6[8],ymm5[9],ymm6[9],ymm5[10],ymm6[10],ymm5[11],ymm6[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm13[4],xmm11[5],xmm13[5],xmm11[6],xmm13[6],xmm11[7],xmm13[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm13, %xmm18
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm7[0],ymm3[0],ymm7[1],ymm3[1],ymm7[2],ymm3[2],ymm7[3],ymm3[3],ymm7[8],ymm3[8],ymm7[9],ymm3[9],ymm7[10],ymm3[10],ymm7[11],ymm3[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm0[4],xmm12[5],xmm0[5],xmm12[6],xmm0[6],xmm12[7],xmm0[7]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %ymm12
; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm12, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm8
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm8[0],ymm0[0],ymm8[1],ymm0[1],ymm8[2],ymm0[2],ymm8[3],ymm0[3],ymm8[8],ymm0[8],ymm8[9],ymm0[9],ymm8[10],ymm0[10],ymm8[11],ymm0[11]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm4
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm8
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0],ymm4[0],ymm8[1],ymm4[1],ymm8[2],ymm4[2],ymm8[3],ymm4[3],ymm8[8],ymm4[8],ymm8[9],ymm4[9],ymm8[10],ymm4[10],ymm8[11],ymm4[11]
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,1,2,3,11,11,11,11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm1[4],ymm12[4],ymm1[5],ymm12[5],ymm1[6],ymm12[6],ymm1[7],ymm12[7],ymm1[12],ymm12[12],ymm1[13],ymm12[13],ymm1[14],ymm12[14],ymm1[15],ymm12[15]
-; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm27, %zmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm26 = [5,6,5,6,5,6,7,7]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512-FCP-NEXT: vpermd %ymm8, %ymm26, %ymm8
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm0[4],ymm12[4],ymm0[5],ymm12[5],ymm0[6],ymm12[6],ymm0[7],ymm12[7],ymm0[12],ymm12[12],ymm0[13],ymm12[13],ymm0[14],ymm12[14],ymm0[15],ymm12[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm27, %zmm8
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm4 = [5,6,5,6,5,6,7,7]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm12 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
+; AVX512-FCP-NEXT: vpermd %ymm12, %ymm4, %ymm12
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm0 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm0
; AVX512-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512-FCP-NEXT: kmovw %eax, %k1
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm0 {%k1}
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [8,21,10,11,20,13,14,23]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm4
-; AVX512-FCP-NEXT: vmovdqa %ymm13, %ymm8
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm29 = [12,1,2,13,4,5,14,7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm29, %ymm0
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm8 {%k1}
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [12,1,2,13,4,5,14,7]
+; AVX512-FCP-NEXT: vmovdqa %ymm8, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm25, %ymm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [8,21,10,11,20,13,14,23]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa %ymm14, %ymm12
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm28, %zmm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [2,2,0,3,10,0,10,11]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,2,0,3,10,0,10,11]
; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm1
; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm0
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm3
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm28, %zmm3
-; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm4
-; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm3
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[8],ymm0[8],ymm3[9],ymm0[9],ymm3[10],ymm0[10],ymm3[11],ymm0[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm4[4],ymm1[4],ymm4[5],ymm1[5],ymm4[6],ymm1[6],ymm4[7],ymm1[7],ymm4[12],ymm1[12],ymm4[13],ymm1[13],ymm4[14],ymm1[14],ymm4[15],ymm1[15]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm0
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
-; AVX512-FCP-NEXT: vpermd %ymm3, %ymm26, %ymm3
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm8
+; AVX512-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm8, %ymm1
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm8[4],ymm2[4],ymm8[5],ymm2[5],ymm8[6],ymm2[6],ymm8[7],ymm2[7],ymm8[12],ymm2[12],ymm8[13],ymm2[13],ymm8[14],ymm2[14],ymm8[15],ymm2[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm0
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[12],ymm6[12],ymm5[13],ymm6[13],ymm5[14],ymm6[14],ymm5[15],ymm6[15]
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm0 {%k1}
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3
-; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm4
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm29, %ymm0
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,2,3]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[8],ymm6[8],ymm5[9],ymm6[9],ymm5[10],ymm6[10],ymm5[11],ymm6[11]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm1
+; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm5 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %ymm5, %ymm25, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm5
+; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm1
; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm3
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm28, %zmm3
-; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm3
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm1, %xmm4
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,2,1,8,9,8,9]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
-; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm31 = [1,0,2,2,1,0,2,2]
-; AVX512-FCP-NEXT: # ymm31 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
-; AVX512-FCP-NEXT: vpermd %ymm3, %ymm31, %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm2
+; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm0
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm5
+; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm5, %xmm1
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,2,1,8,9,8,9]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1
+; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm30 = [1,0,2,2,1,0,2,2]
+; AVX512-FCP-NEXT: # ymm30 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512-FCP-NEXT: vpermd %ymm0, %ymm30, %ymm0
; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm2 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; AVX512-FCP-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512-FCP-NEXT: kmovw %eax, %k2
-; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm4 {%k2}
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm30 = [16,9,10,17,12,13,18,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm2
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm30, %zmm2
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
-; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm16, %ymm4
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm2[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm1 {%k2}
+; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm5
+; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm17, %zmm1
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm26 = [0,1,8,3,4,9,6,7]
+; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm26, %ymm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [0,0,0,0,8,8,0,9]
-; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm17, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm2
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [0,0,0,0,8,8,0,9]
+; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm1
; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm1, %xmm3
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm20
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm18, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm23, %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-FCP-NEXT: vpermd %ymm2, %ymm31, %ymm2
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512-FCP-NEXT: vpbroadcastq %xmm4, %ymm4
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
-; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm0
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm30, %zmm2
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm24
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm16, %ymm3
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm2[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm17, %zmm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm2, %xmm6
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm0
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm5, %xmm1
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm23
+; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm24
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3]
+; AVX512-FCP-NEXT: vpermd %ymm0, %ymm30, %ymm0
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512-FCP-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm1 {%k2}
+; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm5
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm17, %zmm1
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm19
+; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm26, %ymm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm3
+; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm1
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[12],ymm2[12],ymm5[13],ymm2[13],ymm5[14],ymm2[14],ymm5[15],ymm2[15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm21
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm0
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm3[4],ymm7[5],ymm3[5],ymm7[6],ymm3[6],ymm7[7],ymm3[7],ymm7[12],ymm3[12],ymm7[13],ymm3[13],ymm7[14],ymm3[14],ymm7[15],ymm3[15]
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm2 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm3 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[12],ymm0[12],ymm1[13],ymm0[13],ymm1[14],ymm0[14],ymm1[15],ymm0[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm22
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm2
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm10[4],ymm7[4],ymm10[5],ymm7[5],ymm10[6],ymm7[6],ymm10[7],ymm7[7],ymm10[12],ymm7[12],ymm10[13],ymm7[13],ymm10[14],ymm7[14],ymm10[15],ymm7[15]
-; AVX512-FCP-NEXT: vpermd %ymm3, %ymm26, %ymm3
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm5 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[8],ymm4[8],ymm5[9],ymm4[9],ymm5[10],ymm4[10],ymm5[11],ymm4[11]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm2 {%k1}
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm3
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm0
-; AVX512-FCP-NEXT: vmovdqa %ymm8, %ymm5
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm4
-; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm29, %ymm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm3[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm9, %ymm3
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm1
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm1
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm2, %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm21
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm28, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm25, %ymm1
+; AVX512-FCP-NEXT: vmovdqa %ymm12, %ymm5
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm18
+; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm28, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm11
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm11, %ymm1
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm11[4],ymm12[4],ymm11[5],ymm12[5],ymm11[6],ymm12[6],ymm11[7],ymm12[7],ymm11[12],ymm12[12],ymm11[13],ymm12[13],ymm11[14],ymm12[14],ymm11[15],ymm12[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm22
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm29, %zmm2
+; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512-FCP-NEXT: vpermd %ymm3, %ymm26, %ymm3
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm6 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm6[0],ymm4[0],ymm6[1],ymm4[1],ymm6[2],ymm4[2],ymm6[3],ymm4[3],ymm6[8],ymm4[8],ymm6[9],ymm4[9],ymm6[10],ymm4[10],ymm6[11],ymm4[11]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm3, %ymm4
-; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm1, %zmm25
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm29, %ymm1
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm1[0,1,2,3],zmm25[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm5
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm10
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm28, %zmm10
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm6
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm29
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm7
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15]
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm4, %ymm1
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm10
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm25
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
+; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm28, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm25, %zmm25
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm13
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm13
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm14
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm15, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm14, %xmm1
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm4
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512-FCP-NEXT: vpermd %ymm1, %ymm31, %ymm1
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3]
-; AVX512-FCP-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm7 {%k2}
-; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm1
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2d %zmm15, %zmm30, %zmm1
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm15 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512-FCP-NEXT: vpermt2d %ymm15, %ymm16, %ymm7
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm7[0,1,2,3],zmm1[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm15, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm7
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm12
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm1
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm4
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512-FCP-NEXT: vpermd %ymm0, %ymm30, %ymm5
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm3
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm3, %xmm8
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm16, %zmm0
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm8 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm9 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512-FCP-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
+; AVX512-FCP-NEXT: vmovdqa32 %zmm5, %zmm4 {%k2}
+; AVX512-FCP-NEXT: vmovdqa %ymm4, %ymm5
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm9, %zmm17, %zmm4
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm9 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
+; AVX512-FCP-NEXT: vpermt2d %ymm9, %ymm26, %ymm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm16
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512-FCP-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm14[0],xmm2[1],xmm14[1],xmm2[2],xmm14[2],xmm2[3],xmm14[3]
-; AVX512-FCP-NEXT: vpermd %ymm1, %ymm31, %ymm18
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm1 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm14 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3]
-; AVX512-FCP-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm18, %zmm1
-; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm12 {%k2}
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm14
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm14[2,1,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm12, %zmm30
-; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero
-; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm16, %ymm12
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,2,3],zmm30[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm1, %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm1[0,0,2,1,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm17, %zmm12
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [1,1,1,1,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [8,9,20,11,12,21,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm17
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm17
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm14, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [0,9,2,3,8,5,6,11]
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm14, %ymm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm30
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,0,1,0,10,10,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm30, %ymm4
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpsrldq {{.*#+}} xmm9 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
+; AVX512-FCP-NEXT: vpbroadcastq %xmm5, %ymm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vmovdqa32 %zmm4, %zmm0 {%k2}
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm1
+; AVX512-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
+; AVX512-FCP-NEXT: vpermi2d %ymm4, %ymm0, %ymm26
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[2,1,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2d %zmm4, %zmm17, %zmm0
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm9
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm4
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm31, %zmm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm26, %zmm17
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm31, %zmm4
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm11[0],ymm12[0],ymm11[1],ymm12[1],ymm11[2],ymm12[2],ymm11[3],ymm12[3],ymm11[8],ymm12[8],ymm11[9],ymm12[9],ymm11[10],ymm12[10],ymm11[11],ymm12[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm12 = [1,1,1,1,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm3
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm26 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm2 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm2 {%k1}
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm3
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [0,9,2,3,8,5,6,11]
+; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm27, %ymm1
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm30 = [8,9,20,11,12,21,14,15]
+; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm30, %zmm2
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm0, %xmm3
+; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm31
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,0,10,10,0]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm28, %zmm3
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm5 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm7 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm1 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm9, %zmm1
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm5 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa32 %zmm1, %zmm5 {%k1}
-; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm28
-; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm18, %zmm28
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm0
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm7 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm7 {%k1}
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm1
-; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm14, %ymm5
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm1 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm13
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm9, %zmm0
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm29
+; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm27, %ymm29
; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
; AVX512-FCP-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm20
-; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm20
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm14, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm13
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm13[0],ymm0[1],ymm13[1],ymm0[2],ymm13[2],ymm0[3],ymm13[3],ymm0[8],ymm13[8],ymm0[9],ymm13[9],ymm0[10],ymm13[10],ymm0[11],ymm13[11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm13
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm13[4],xmm11[5],xmm13[5],xmm11[6],xmm13[6],xmm11[7],xmm13[7]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm11
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm0 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa32 %zmm11, %zmm0 {%k1}
-; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm9
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512-FCP-NEXT: vpermi2d %zmm9, %zmm0, %zmm18
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpermt2d %ymm6, %ymm14, %ymm0
-; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm11
+; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm7
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm23, %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm6
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm10
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa32 %zmm10, %zmm0 {%k1}
+; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm6, %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512-FCP-NEXT: vpermt2d %ymm10, %ymm27, %ymm19
+; AVX512-FCP-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm10 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %zmm10, %zmm30, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm10
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm6[0],ymm10[0],ymm6[1],ymm10[1],ymm6[2],ymm10[2],ymm6[3],ymm10[3],ymm6[8],ymm10[8],ymm6[9],ymm10[9],ymm6[10],ymm10[10],ymm6[11],ymm10[11]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
+; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm12, %zmm14
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm10 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa32 %zmm14, %zmm10 {%k1}
; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm27, %zmm8
-; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm27, %zmm9
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm7
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm27, %zmm6
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm18[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm6
+; AVX512-FCP-NEXT: vpermi2d %ymm8, %ymm10, %ymm27
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm6
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2d %zmm8, %zmm30, %zmm10
+; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm8 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm14
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm28, %zmm11
+; AVX512-FCP-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm8 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm12
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm28, %zmm12
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm9, %xmm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm28, %zmm6
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm27, %zmm8
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm8, %zmm9, %zmm6
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm20[0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm9
-; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 448(%rax)
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm28[0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm8
-; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 640(%rax)
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm17[0,1,2,3]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm3
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm19, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm9, %zmm12
+; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 448(%rax)
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm29, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm9, %zmm11
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 640(%rax)
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm9, %zmm3
; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm16, %zmm0, %zmm12
-; AVX512-FCP-NEXT: vmovdqa64 %zmm12, (%rax)
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm26, %zmm0, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm17, %zmm0, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm16, %zmm0, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm25, %zmm1, %zmm10
-; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm25, %zmm1, %zmm13
+; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 128(%rax)
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 320(%rax)
@@ -9244,15 +9201,15 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
; AVX512-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 704(%rax)
-; AVX512-FCP-NEXT: addq $1256, %rsp # imm = 0x4E8
+; AVX512-FCP-NEXT: addq $1240, %rsp # imm = 0x4D8
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i16_stride6_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $936, %rsp # imm = 0x3A8
+; AVX512DQ-NEXT: subq $584, %rsp # imm = 0x248
; AVX512DQ-NEXT: vmovdqa 96(%rcx), %xmm0
-; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX512DQ-NEXT: vmovdqa 96(%rdx), %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
@@ -9262,7 +9219,7 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm0, %zmm1
; AVX512DQ-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512DQ-NEXT: vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm4[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vmovdqa 96(%rdi), %xmm5
@@ -9282,22 +9239,12 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX512DQ-NEXT: vpbroadcastq %xmm3, %ymm3
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 96(%r9), %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm5
; AVX512DQ-NEXT: vmovdqa 64(%rcx), %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm6
@@ -9308,57 +9255,47 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm0, %zmm1
-; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm5
-; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm5[0,1,2,1]
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm4
+; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm7
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
-; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm7
-; AVX512DQ-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[0,1,2,1]
+; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm8
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm8[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm29
+; AVX512DQ-NEXT: vmovdqa64 %xmm7, %xmm30
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[0,1,0,1]
; AVX512DQ-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa 64(%r8), %xmm5
-; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vmovdqa 64(%r8), %xmm7
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm7[2,1,3,3,4,5,6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
+; AVX512DQ-NEXT: vmovdqa64 %xmm7, %xmm31
; AVX512DQ-NEXT: vpbroadcastq %xmm3, %ymm3
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm7
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 64(%r9), %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm0, %zmm2
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm5
-; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
+; AVX512DQ-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm1 {%k1}
@@ -9368,39 +9305,30 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm3
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm2
; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
; AVX512DQ-NEXT: vpbroadcastq %xmm3, %ymm3
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm3
; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm4
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm31
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm19
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm27
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm28
; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm0, %zmm2
; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm3
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm3[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm15
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm15[0,1,2,1]
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[0,1,2,1]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm15[0],xmm3[0],xmm15[1],xmm3[1],xmm15[2],xmm3[2],xmm15[3],xmm3[3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm18
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm20
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm21
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[0,1,0,1]
; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
@@ -9409,12 +9337,11 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm26
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm18
; AVX512DQ-NEXT: vpbroadcastq %xmm2, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 96(%rsi), %ymm2
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
@@ -9425,34 +9352,32 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm28
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm26
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa 96(%rcx), %ymm4
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa 96(%rdx), %ymm11
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 96(%rcx), %ymm12
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 96(%rdx), %ymm9
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm11[4],ymm4[4],ymm11[5],ymm4[5],ymm11[6],ymm4[6],ymm11[7],ymm4[7],ymm11[12],ymm4[12],ymm11[13],ymm4[13],ymm11[14],ymm4[14],ymm11[15],ymm4[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm21
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm9[4],ymm12[4],ymm9[5],ymm12[5],ymm9[6],ymm12[6],ymm9[7],ymm12[7],ymm9[12],ymm12[12],ymm9[13],ymm12[13],ymm9[14],ymm12[14],ymm9[15],ymm12[15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-NEXT: kmovw %eax, %k1
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
-; AVX512DQ-NEXT: vmovdqa 96(%r8), %ymm9
+; AVX512DQ-NEXT: vmovdqa 96(%r8), %ymm6
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm9, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm6, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 64(%rsi), %ymm2
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
@@ -9463,50 +9388,49 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm16
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm17
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm12
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm12[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm12[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm10
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm7
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm5
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[12],ymm12[12],ymm10[13],ymm12[13],ymm10[14],ymm12[14],ymm10[15],ymm12[15]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm5[4],ymm7[4],ymm5[5],ymm7[5],ymm5[6],ymm7[6],ymm5[7],ymm7[7],ymm5[12],ymm7[12],ymm5[13],ymm7[13],ymm5[14],ymm7[14],ymm5[15],ymm7[15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
-; AVX512DQ-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm3, %ymm1
+; AVX512DQ-NEXT: vmovdqa 64(%r8), %ymm8
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm8, %ymm1
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm8[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm14
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm14[4],ymm2[4],ymm14[5],ymm2[5],ymm14[6],ymm2[6],ymm14[7],ymm2[7],ymm14[12],ymm2[12],ymm14[13],ymm2[13],ymm14[14],ymm2[14],ymm14[15],ymm2[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm6
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm11
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm2 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm10
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm3 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[12],ymm7[12],ymm6[13],ymm7[13],ymm6[14],ymm7[14],ymm6[15],ymm7[15]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[12],ymm11[12],ymm10[13],ymm11[13],ymm10[14],ymm11[14],ymm10[15],ymm11[15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
@@ -9516,252 +9440,257 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm19
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm3
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm17 = ymm0[2,1,2,3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm5[4],ymm3[5],ymm5[5],ymm3[6],ymm5[6],ymm3[7],ymm5[7],ymm3[12],ymm5[12],ymm3[13],ymm5[13],ymm3[14],ymm5[14],ymm3[15],ymm5[15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm16 = ymm0[3,3,3,3]
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm1 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm8 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,1,2,3]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm23 = ymm0[3,3,3,3]
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm1 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512DQ-NEXT: vpsrldq {{.*#+}} ymm15 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm15[0],ymm1[0],ymm15[1],ymm1[1],ymm15[2],ymm1[2],ymm15[3],ymm1[3],ymm15[8],ymm1[8],ymm15[9],ymm1[9],ymm15[10],ymm1[10],ymm15[11],ymm1[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm25 = ymm1[2,2,2,2]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm15 = ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[12],ymm2[12],ymm0[13],ymm2[13],ymm0[14],ymm2[14],ymm0[15],ymm2[15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[1,2,3,3,5,6,7,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm1
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm15, %zmm25, %zmm1 {%k1}
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm15, %ymm14
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm13
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3],ymm14[4],ymm13[5,6],ymm14[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm14[0],ymm1[1,2],ymm14[3],ymm1[4,5],ymm14[6],ymm1[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm1, %zmm23
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm13
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm24 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
+; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm24, %zmm0
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm2
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm2[4],ymm4[4],ymm2[5],ymm4[5],ymm2[6],ymm4[6],ymm2[7],ymm4[7],ymm2[12],ymm4[12],ymm2[13],ymm4[13],ymm2[14],ymm4[14],ymm2[15],ymm4[15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[1,2,3,3,5,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm16, %zmm17, %zmm0
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm0 {%k1}
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm1, %ymm8
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm13
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0],ymm8[1],ymm13[2,3],ymm8[4],ymm13[5,6],ymm8[7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm13[0],ymm0[1,2],ymm13[3],ymm0[4,5],ymm13[6],ymm0[7]
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[0,1,2,3],zmm8[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm8
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm8[0,0,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpbroadcastq %xmm0, %ymm30
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm19, %xmm13
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm13 = xmm8[0,2,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,4,4,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm13[0,0,2,1]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [17,18,17,18,0,0,19,19,5,4,2,2,5,4,6,6]
-; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm4, %zmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm0
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
-; AVX512DQ-NEXT: vmovdqa 96(%r9), %ymm15
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm15[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm23 = ymm13[2,2,2,3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm5[0],ymm3[1],ymm5[1],ymm3[2],ymm5[2],ymm3[3],ymm5[3],ymm3[8],ymm5[8],ymm3[9],ymm5[9],ymm3[10],ymm5[10],ymm3[11],ymm5[11]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vmovdqa64 %ymm15, %ymm16
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm5[2,1,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm0, %zmm2 {%k1}
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm2, %ymm1
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
-; AVX512DQ-NEXT: vmovdqa 64(%r9), %ymm0
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vmovdqa %ymm0, %ymm15
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm17 = ymm5[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm2[0,1,2,3],zmm3[4,5,6,7]
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm2
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm4
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm25
; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[8],ymm0[8],ymm11[9],ymm0[9],ymm11[10],ymm0[10],ymm11[11],ymm0[11]
-; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm4, %zmm11
+; AVX512DQ-NEXT: vpunpckhwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm12[0],ymm9[1],ymm12[1],ymm9[2],ymm12[2],ymm9[3],ymm12[3],ymm9[8],ymm12[8],ymm9[9],ymm12[9],ymm9[10],ymm12[10],ymm9[11],ymm12[11]
+; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm24, %zmm1
; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd (%rsp), %xmm0, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
+; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm2, %zmm0, %zmm1 {%k1}
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm2
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm2, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm4, %xmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm26
+; AVX512DQ-NEXT: vmovdqa 96(%r9), %xmm6
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm6[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm7[0],ymm5[1],ymm7[1],ymm5[2],ymm7[2],ymm5[3],ymm7[3],ymm5[8],ymm7[8],ymm5[9],ymm7[9],ymm5[10],ymm7[10],ymm5[11],ymm7[11]
+; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm24, %zmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm28, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm27, %ymm3
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[8],ymm3[8],ymm0[9],ymm3[9],ymm0[10],ymm3[10],ymm0[11],ymm3[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm11 {%k1}
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm4
+; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm5
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm16
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm16 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm4[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm1, %zmm2, %zmm3 {%k1}
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm3, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm8[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm11, %ymm3
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm9 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[8],ymm12[8],ymm10[9],ymm12[9],ymm10[10],ymm12[10],ymm10[11],ymm12[11]
-; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm4, %zmm13
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm24, %ymm3
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[8],ymm3[8],ymm0[9],ymm3[9],ymm0[10],ymm3[10],ymm0[11],ymm3[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm10 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[8],ymm7[8],ymm6[9],ymm7[9],ymm6[10],ymm7[10],ymm6[11],ymm7[11]
-; AVX512DQ-NEXT: vpermt2d %zmm10, %zmm4, %zmm5
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm4 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm14[0],ymm0[0],ymm14[1],ymm0[1],ymm14[2],ymm0[2],ymm14[3],ymm0[3],ymm14[8],ymm0[8],ymm14[9],ymm0[9],ymm14[10],ymm0[10],ymm14[11],ymm0[11]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm13 {%k1}
-; AVX512DQ-NEXT: vinserti32x8 $1, %ymm6, %zmm4, %zmm5 {%k1}
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0],ymm7[1],ymm11[2,3],ymm7[4],ymm11[5,6],ymm7[7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm15[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vmovdqa %ymm15, %ymm6
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm14
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm9, %xmm0, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm9, %xmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6],ymm2[7]
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm8
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm27
+; AVX512DQ-NEXT: vmovdqa 64(%r9), %xmm9
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm9[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm9[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm8[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm2, %ymm0
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm8[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm2[0,0,2,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm14[0,0,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpbroadcastq %xmm2, %ymm29
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm3 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm3 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm14[0,2,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm2[0,0,2,1]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[8],ymm11[8],ymm10[9],ymm11[9],ymm10[10],ymm11[10],ymm10[11],ymm11[11]
+; AVX512DQ-NEXT: vmovdqa 96(%r9), %ymm11
+; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm24, %zmm2
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm3 = ymm11[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm3[2,2,2,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm11[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm4[2,1,2,3]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm10 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm10 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
+; AVX512DQ-NEXT: vmovdqa 64(%r9), %ymm4
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm7 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm7 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[8],mem[8],ymm3[9],mem[9],ymm3[10],mem[10],ymm3[11],mem[11]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm13[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[1,1,1,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-NEXT: vinserti32x8 $1, %ymm7, %zmm10, %zmm2 {%k1}
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[2,3,2,3,6,7,6,7]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm21 = ymm7[2,1,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm10[0,1,2,3],zmm9[4,5,6,7]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm13, %ymm10
-; AVX512DQ-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm11 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm11 = ymm10[0,1],ymm11[2],ymm10[3,4],ymm11[5],ymm10[6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm2, %ymm10
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm15 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm15 = ymm10[0,1],ymm15[2],ymm10[3,4],ymm15[5],ymm10[6,7]
; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm10
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm12
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0],ymm12[1],ymm13[2,3],ymm12[4],ymm13[5,6],ymm12[7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm10[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm13[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm12[0,1,2,3],zmm11[4,5,6,7]
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm5, %ymm12
-; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm0
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm10[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm1
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm9
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1],ymm5[2,3],ymm1[4],ymm5[5,6],ymm1[7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm9[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm5 = ymm10[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
+; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6],ymm3[7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm1 # 64-byte Folded Reload
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm12
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm12[4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm15, %zmm2, %zmm2
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm15
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm15[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm13 = ymm15[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,1,2,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,2]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,3,2,3]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,2,1,4,5,6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512DQ-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm15 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm7 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm13 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm25
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm25 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm18, %zmm23, %zmm6
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm21, %zmm17, %zmm16
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm16 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm24, %zmm14
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm14 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm5
-; AVX512DQ-NEXT: vpternlogd $184, %zmm29, %zmm18, %zmm5
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm19, %zmm9, %zmm8
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512DQ-NEXT: vpternlogd $184, %zmm20, %zmm9, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm7
-; AVX512DQ-NEXT: vpternlogd $184, %zmm22, %zmm9, %zmm7
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm13, %zmm10
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm9, %zmm10
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm12 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm30, %zmm29, %zmm17
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm17 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm24, %zmm22
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm22 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm7
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm7 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm5, %zmm3
+; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpternlogd $184, %zmm23, %zmm24, %zmm0
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm5
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm25, %zmm13, %zmm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm6, %zmm6
+; AVX512DQ-NEXT: vpternlogd $184, %zmm26, %zmm13, %zmm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm9, %zmm4
+; AVX512DQ-NEXT: vpternlogd $184, %zmm27, %zmm13, %zmm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm13, %zmm8
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm10, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm7, 448(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm0, 640(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm25, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm3, 384(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 576(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm5, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm16, 512(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm6, 704(%rax)
-; AVX512DQ-NEXT: addq $936, %rsp # imm = 0x3A8
+; AVX512DQ-NEXT: vmovdqa64 %zmm8, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm4, 448(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm6, 640(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm5, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm17, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm12, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm16, 576(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm3, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm7, 512(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm22, 704(%rax)
+; AVX512DQ-NEXT: addq $584, %rsp # imm = 0x248
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i16_stride6_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $1224, %rsp # imm = 0x4C8
+; AVX512DQ-FCP-NEXT: subq $1176, %rsp # imm = 0x498
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm0
@@ -9771,8 +9700,8 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm4
; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm5
@@ -9782,95 +9711,95 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %ymm6
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[8],ymm2[8],ymm6[9],ymm2[9],ymm6[10],ymm2[10],ymm6[11],ymm2[11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm30
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %ymm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm7[0],ymm2[0],ymm7[1],ymm2[1],ymm7[2],ymm2[2],ymm7[3],ymm2[3],ymm7[8],ymm2[8],ymm7[9],ymm2[9],ymm7[10],ymm2[10],ymm7[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm29
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm31
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %xmm2
; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm8
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm8[0],ymm1[0],ymm8[1],ymm1[1],ymm8[2],ymm1[2],ymm8[3],ymm1[3],ymm8[8],ymm1[8],ymm8[9],ymm1[9],ymm8[10],ymm1[10],ymm8[11],ymm1[11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm8, %ymm28
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm17
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm8, %ymm30
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm16
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [1,1,1,1,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm3
; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
-; AVX512DQ-FCP-NEXT: vpermd %zmm0, %zmm20, %zmm22
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm19 = [1,2,1,2,0,0,3,3,13,12,10,10,13,12,14,14]
+; AVX512DQ-FCP-NEXT: vpermd %zmm0, %zmm19, %zmm21
; AVX512DQ-FCP-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FCP-NEXT: kmovw %eax, %k1
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm22 {%k1}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm25 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm21 {%k1}
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm24 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm24, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [8,9,20,11,12,21,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm25, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm19 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm19, %ymm22
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,0,0,1,0,10,10,0]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm21
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,0,0,1,0,10,10,0]
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm1
; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %ymm9
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %ymm5
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm9[0],ymm5[1],ymm9[1],ymm5[2],ymm9[2],ymm5[3],ymm9[3],ymm5[8],ymm9[8],ymm5[9],ymm9[9],ymm5[10],ymm9[10],ymm5[11],ymm9[11]
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm14
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm21
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
-; AVX512DQ-FCP-NEXT: vpermd %zmm6, %zmm20, %zmm26
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm4, %zmm26 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm15
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm23
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm17, %zmm4
+; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm19, %zmm25
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm4, %zmm25 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm1, %xmm2
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm2, %ymm24, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm10
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm25, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm24
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm19, %ymm26
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm18, %zmm25
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm11
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm4 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm22, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
@@ -9880,332 +9809,329 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm12
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm18, %zmm6
-; AVX512DQ-FCP-NEXT: vpermd %zmm17, %zmm20, %zmm17
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm6, %zmm17 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm6
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm25, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm17, %zmm6
+; AVX512DQ-FCP-NEXT: vpermd %zmm20, %zmm19, %zmm28
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm6, %zmm28 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm20
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm7
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm7, %ymm19, %ymm17
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm24, %ymm20
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm6
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm18, %zmm28
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm1
; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %ymm7
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm3
; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[8],ymm2[8],ymm0[9],ymm2[9],ymm0[10],ymm2[10],ymm0[11],ymm2[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm18, %zmm0
-; AVX512DQ-FCP-NEXT: vpermd %zmm16, %zmm20, %zmm20
-; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm17, %zmm0
+; AVX512DQ-FCP-NEXT: vpermd %zmm16, %zmm19, %zmm19
+; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm0, %zmm19 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm0, %ymm19, %ymm24
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm20, %zmm25
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm19, %ymm20
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm18, %zmm19
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm0, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm9, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm1
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm5[4],ymm9[4],ymm5[5],ymm9[5],ymm5[6],ymm9[6],ymm5[7],ymm9[7],ymm5[12],ymm9[12],ymm5[13],ymm9[13],ymm5[14],ymm9[14],ymm5[15],ymm9[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,1,2,3,11,11,11,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm8
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm27 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm8
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm26 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm27, %ymm0
+; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm26, %ymm0
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm5 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm5[0],ymm1[0],ymm5[1],ymm1[1],ymm5[2],ymm1[2],ymm5[3],ymm1[3],ymm5[8],ymm1[8],ymm5[9],ymm1[9],ymm5[10],ymm1[10],ymm5[11],ymm1[11]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm8 {%k1}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm22 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm22, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31]
; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm9
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm30 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm30, %ymm8
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm0[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm10
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm29 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm29, %zmm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm5
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm11, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm8
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [2,2,0,3,10,0,10,11]
; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm12, %ymm2
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm4
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm27, %ymm2
+; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm26, %ymm2
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm3 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm0 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm9, %ymm10
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm23, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm2
; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm3 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm30, %ymm0
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm22, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm29, %zmm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm2
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm4, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm14, %xmm3
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,0,2,1,8,9,8,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm3
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm8, %ymm12
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm23, %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm4, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm15, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm4[0],xmm15[1],xmm4[1],xmm15[2],xmm4[2],xmm15[3],xmm4[3]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,2,1,8,9,8,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm2
+; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm30 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FCP-NEXT: # ymm30 = mem[0,1,2,3,0,1,2,3]
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm30, %ymm0
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm5, %ymm5
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm4, %ymm4
; AVX512DQ-FCP-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FCP-NEXT: kmovw %eax, %k2
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm5, %zmm4, %zmm3 {%k2}
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm28 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm6
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm4
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm24[0],zero,xmm24[1],zero,xmm24[2],zero,xmm24[3],zero
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm5, %ymm16, %ymm3
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm4, %zmm0, %zmm2 {%k2}
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm5
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm0, %ymm5
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm5, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, %xmm8
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,0,0,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm18, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm5, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm5, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, %xmm9
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,0,0,8,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm4
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm4
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm7, %xmm5
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm18, %zmm5
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpermd %ymm3, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm5, %zmm3, %zmm4 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm5, %ymm16, %ymm4
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm4[0,1,2,3],zmm3[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm4, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm8, %xmm21
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm5
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm3[0],ymm5[1],ymm3[1],ymm5[2],ymm3[2],ymm5[3],ymm3[3],ymm5[8],ymm3[8],ymm5[9],ymm3[9],ymm5[10],ymm3[10],ymm5[11],ymm3[11]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm6[4],ymm8[4],ymm6[5],ymm8[5],ymm6[6],ymm8[6],ymm6[7],ymm8[7],ymm6[12],ymm8[12],ymm6[13],ymm8[13],ymm6[14],ymm8[14],ymm6[15],ymm8[15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm29, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[12],ymm6[12],ymm8[13],ymm6[13],ymm8[14],ymm6[14],ymm8[15],ymm6[15]
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm8 = ymm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm8[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[8],ymm6[8],ymm8[9],ymm6[9],ymm8[10],ymm6[10],ymm8[11],ymm6[11]
-; AVX512DQ-FCP-NEXT: vpermd %ymm5, %ymm27, %ymm5
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm5, %zmm6, %zmm3 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm10, %ymm14
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm23, %zmm6
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm5 = ymm8[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm5, %ymm30, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm13, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm10, %ymm8
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm8[0],ymm5[0],ymm8[1],ymm5[1],ymm8[2],ymm5[2],ymm8[3],ymm5[3],ymm8[8],ymm5[8],ymm8[9],ymm5[9],ymm8[10],ymm5[10],ymm8[11],ymm5[11]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm10[4],ymm13[4],ymm10[5],ymm13[5],ymm10[6],ymm13[6],ymm10[7],ymm13[7],ymm10[12],ymm13[12],ymm10[13],ymm13[13],ymm10[14],ymm13[14],ymm10[15],ymm13[15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm29, %zmm5
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm3[0,1,2,3],zmm6[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm3, %ymm8
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm11, %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm31, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm7 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm30, %ymm2
+; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm6, %zmm2, %zmm5 {%k2}
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm7[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm17, %zmm5
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm6, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm23
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm5, %xmm2
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm9, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0],ymm2[0],ymm6[1],ymm2[1],ymm6[2],ymm2[2],ymm6[3],ymm2[3],ymm6[8],ymm2[8],ymm6[9],ymm2[9],ymm6[10],ymm2[10],ymm6[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm7[4],ymm9[4],ymm7[5],ymm9[5],ymm7[6],ymm9[6],ymm7[7],ymm9[7],ymm7[12],ymm9[12],ymm7[13],ymm9[13],ymm7[14],ymm9[14],ymm7[15],ymm9[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm9[4],ymm7[4],ymm9[5],ymm7[5],ymm9[6],ymm7[6],ymm9[7],ymm7[7],ymm9[12],ymm7[12],ymm9[13],ymm7[13],ymm9[14],ymm7[14],ymm9[15],ymm7[15]
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm7 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[2],ymm7[2],ymm9[3],ymm7[3],ymm9[8],ymm7[8],ymm9[9],ymm7[9],ymm9[10],ymm7[10],ymm9[11],ymm7[11]
+; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm26, %ymm2
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm2, %zmm7, %zmm6 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm2, %ymm22, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm10, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm2
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm29, %zmm6
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm11[4],ymm6[4],ymm11[5],ymm6[5],ymm11[6],ymm6[6],ymm11[7],ymm6[7],ymm11[12],ymm6[12],ymm11[13],ymm6[13],ymm11[14],ymm6[14],ymm11[15],ymm6[15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm27, %ymm8
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm10 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm11 = ymm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm11[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[8],ymm10[8],ymm11[9],ymm10[9],ymm11[10],ymm10[10],ymm11[11],ymm10[11]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm8, %zmm10, %zmm5 {%k1}
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm8
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm8, %zmm5, %zmm23
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm8 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm8, %ymm30, %ymm5
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm11, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm10, %ymm9
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm9[0],ymm2[0],ymm9[1],ymm2[1],ymm9[2],ymm2[2],ymm9[3],ymm2[3],ymm9[8],ymm2[8],ymm9[9],ymm2[9],ymm9[10],ymm2[10],ymm9[11],ymm2[11]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[12],ymm11[12],ymm10[13],ymm11[13],ymm10[14],ymm11[14],ymm10[15],ymm11[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm27
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm6, %ymm10
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm8
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm31, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm13, %xmm11
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm19, %zmm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm0, %ymm11
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm13 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm14 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm12, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm31, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15]
+; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm26, %ymm7
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm11 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} ymm12 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm7, %zmm11, %zmm2 {%k1}
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm7 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm7, %ymm2, %ymm22
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm9, %ymm7
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm29, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm31, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsp), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm12
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm18, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512DQ-FCP-NEXT: vpermd %ymm12, %ymm30, %ymm12
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm13 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm14 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm13, %ymm13
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm13, %zmm11, %zmm10 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsp), %xmm15 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm13 = xmm15[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm28, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm14, %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm9
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm6[0],xmm14[0],xmm6[1],xmm14[1],xmm6[2],xmm14[2],xmm6[3],xmm14[3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm13
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm9 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3]
-; AVX512DQ-FCP-NEXT: vpermd %ymm14, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm14 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm14, %ymm14
-; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm14, %zmm0, %zmm13 {%k2}
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm13, %zmm28
-; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm9, %ymm16, %ymm10
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm16, %ymm13
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm23[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm11[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm6, %xmm9
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm28[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm6, %xmm11
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm13 = xmm6[0,0,2,1,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm18, %zmm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm9, %zmm11, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm14
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm18, %zmm8
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm13, %zmm12, %zmm11 {%k2}
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm12 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm13 = xmm9[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm11, %ymm14
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm17, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vpermd %ymm13, %ymm30, %ymm13
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm15 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpsrldq {{.*#+}} xmm10 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3]
+; AVX512DQ-FCP-NEXT: vpbroadcastq %xmm10, %ymm10
+; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm10, %zmm13, %zmm8 {%k2}
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm12, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpmovzxwd {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
+; AVX512DQ-FCP-NEXT: vpermi2d %ymm10, %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm9[2,1,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm10, %zmm17, %zmm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm22, %zmm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm14, %zmm10
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm9, %xmm11
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm9[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm16, %zmm12
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm9[0,0,2,1,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm16, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm8, %zmm11
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, (%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm5, %zmm11, %zmm10
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm29, %zmm5, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 320(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm20[0,1,2,3],zmm25[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm17[0,1,2,3],mem[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 448(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm12, %zmm11, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 384(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 576(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 512(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, (%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm10, %zmm8, %zmm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm0, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm27, %zmm0, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 320(%rax)
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm24, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm6, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 256(%rax)
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm28, %zmm20, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm6, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 448(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm23, %zmm8, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 384(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 576(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 512(%rax)
+; AVX512DQ-FCP-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 704(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm26[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 640(%rax)
-; AVX512DQ-FCP-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm0 = zmm22[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-FCP-NEXT: addq $1224, %rsp # imm = 0x4C8
+; AVX512DQ-FCP-NEXT: addq $1176, %rsp # imm = 0x498
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index 208ee607909ed..dc362d729fcd3 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
@@ -2850,8 +2850,8 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-NEXT: vporq %ymm1, %ymm4, %ymm17
; AVX512-NEXT: vmovdqa (%rdi), %xmm1
; AVX512-NEXT: vmovdqa (%rsi), %xmm4
-; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[16,17,u,u,u,u]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[16,17,u,u],zero,zero
+; AVX512-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u]
; AVX512-NEXT: vporq %ymm5, %ymm10, %ymm19
; AVX512-NEXT: vpbroadcastd 8(%rax), %ymm5
; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
@@ -2982,8 +2982,8 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,14,15],zero,zero,ymm4[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm4[u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[16,17,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17,u,u],zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm18
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
@@ -3112,8 +3112,8 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-NEXT: vporq %ymm1, %ymm4, %ymm17
; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm1
; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm4
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[16,17,u,u,u,u]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm5 = ymm3[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm3[16,17,u,u],zero,zero
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u]
; AVX512DQ-NEXT: vporq %ymm5, %ymm10, %ymm19
; AVX512DQ-NEXT: vpbroadcastd 8(%rax), %ymm5
; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm5
@@ -3244,8 +3244,8 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,14,15],zero,zero,ymm4[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm4[u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[16,17,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15,u,u],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17,u,u],zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm18
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
@@ -5893,1237 +5893,1257 @@ define void @store_i16_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride7_vf32:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $632, %rsp # imm = 0x278
+; AVX512-NEXT: subq $680, %rsp # imm = 0x2A8
; AVX512-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm9, %ymm1, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-NEXT: vpshufb %ymm13, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm29
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm13, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm27
+; AVX512-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512-NEXT: vpshufb %ymm14, %ymm8, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-NEXT: vmovdqa (%rsi), %ymm9
; AVX512-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm12, %ymm1, %ymm0
-; AVX512-NEXT: vmovdqa %ymm1, %ymm15
-; AVX512-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm14 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-NEXT: vpshufb %ymm14, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512-NEXT: vpshufb %ymm12, %ymm9, %ymm0
+; AVX512-NEXT: vmovdqa (%rdi), %ymm11
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm15 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512-NEXT: vpshufb %ymm15, %ymm11, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %ymm10
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm1, %ymm10, %ymm2
-; AVX512-NEXT: vmovdqa (%r8), %ymm11
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-NEXT: vpshufb %ymm0, %ymm11, %ymm3
+; AVX512-NEXT: vmovdqa (%r9), %ymm1
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm0, %ymm1, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512-NEXT: vmovdqa (%r8), %ymm4
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm17
; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r9), %xmm3
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm7 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm7
-; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,6]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm10
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3]
+; AVX512-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,6]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm4 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
-; AVX512-NEXT: vpermi2d %zmm2, %zmm3, %zmm4
+; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm4
; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm6
-; AVX512-NEXT: vpshufb %ymm12, %ymm6, %ymm2
-; AVX512-NEXT: vpshufb %ymm14, %ymm3, %ymm4
-; AVX512-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm10
+; AVX512-NEXT: vpshufb %ymm12, %ymm10, %ymm3
+; AVX512-NEXT: vpshufb %ymm15, %ymm2, %ymm4
+; AVX512-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa 32(%rcx), %ymm12
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm7
-; AVX512-NEXT: vpshufb %ymm9, %ymm12, %ymm2
-; AVX512-NEXT: vpshufb %ymm13, %ymm7, %ymm4
-; AVX512-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm15
+; AVX512-NEXT: vpshufb %ymm13, %ymm12, %ymm3
+; AVX512-NEXT: vpshufb %ymm14, %ymm15, %ymm4
+; AVX512-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa 32(%r9), %ymm13
; AVX512-NEXT: vmovdqa 32(%r8), %ymm14
-; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512-NEXT: vpshufb %ymm0, %ymm14, %ymm0
-; AVX512-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm2 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,17,17,0,0,0,0,0,1,2,0,0,3]
-; AVX512-NEXT: vpermi2d %zmm2, %zmm1, %zmm4
-; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-NEXT: vmovdqa (%r8), %xmm2
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,5,7,6]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm9 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
-; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm9
-; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vprold $16, %ymm10, %ymm4
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm9[0,1,2],ymm5[3],ymm9[4,5],ymm5[6],ymm9[7,8,9,10],ymm5[11],ymm9[12,13],ymm5[14],ymm9[15]
-; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,2,3,3,10,9,11,10]
-; AVX512-NEXT: vpermi2q %zmm4, %zmm5, %zmm27
+; AVX512-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm1 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm4
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3],xmm4[4],xmm5[5,6],xmm4[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm26 = [0,16,0,0,17,17,0,0,0,0,0,1,2,0,0,3]
+; AVX512-NEXT: vpermi2d %zmm5, %zmm4, %zmm26
+; AVX512-NEXT: vmovdqa (%r9), %xmm4
+; AVX512-NEXT: vmovdqa (%r8), %xmm5
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,4,5,7,6]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
+; AVX512-NEXT: vpermi2d %zmm7, %zmm6, %zmm25
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm31
-; AVX512-NEXT: vmovdqa (%rax), %ymm4
-; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT: vpbroadcastd 8(%rax), %ymm0
; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-NEXT: vpshufb %ymm5, %ymm4, %ymm4
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm28
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512-NEXT: vmovdqa (%rax), %ymm3
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm7 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-NEXT: vpshufb %ymm7, %ymm3, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm7, %ymm21
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm11 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512-NEXT: vprold $16, %ymm13, %ymm0
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm13[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm28
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm29
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm30
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512-NEXT: vprold $16, %ymm13, %ymm4
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm14[1,2,2,3,5,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm13[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm27, %ymm3
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT: vmovdqa64 %ymm16, %ymm4
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm29[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512-NEXT: vmovdqa %ymm15, %ymm8
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512-NEXT: vprold $16, %xmm9, %xmm1
-; AVX512-NEXT: vpshufd {{.*#+}} xmm15 = xmm0[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0,1],xmm1[2],xmm15[3,4],xmm1[5],xmm15[6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512-NEXT: vprold $16, %xmm5, %xmm6
+; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2],xmm7[3,4],xmm6[5],xmm7[6,7]
+; AVX512-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512-NEXT: vpshufb %xmm1, %xmm4, %xmm1
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm7[0],xmm1[1],xmm7[2,3],xmm1[4],xmm7[5,6],xmm1[7]
; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
-; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm9
-; AVX512-NEXT: vmovdqa64 %xmm20, %xmm0
-; AVX512-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm15
-; AVX512-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0],xmm0[1],xmm10[2,3],xmm0[4],xmm10[5,6],xmm0[7]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm26
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
-; AVX512-NEXT: vmovdqa64 %xmm9, %xmm25
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1],ymm3[2],ymm6[3,4],ymm3[5],ymm6[6,7,8,9],ymm3[10],ymm6[11,12],ymm3[13],ymm6[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm24
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm6 = ymm12[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm6 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1,2],ymm3[3],ymm6[4,5],ymm3[6],ymm6[7,8,9,10],ymm3[11],ymm6[12,13],ymm3[14],ymm6[15]
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm21
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
-; AVX512-NEXT: vmovdqa 32(%rax), %ymm6
-; AVX512-NEXT: vpermd %zmm6, %zmm3, %zmm3
-; AVX512-NEXT: vmovdqa64 %ymm28, %ymm7
-; AVX512-NEXT: vpshufb %ymm7, %ymm6, %ymm10
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,2,3]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm6
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-NEXT: vpshufb %xmm7, %xmm2, %xmm10
-; AVX512-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqa (%rdi), %xmm13
-; AVX512-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
-; AVX512-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512-NEXT: vprold $16, %xmm14, %xmm14
-; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm16
-; AVX512-NEXT: vmovdqa %ymm4, %ymm2
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vmovdqa64 %ymm29, %ymm12
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm29[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm14 = ymm9[0,1],ymm14[2],ymm9[3,4],ymm14[5],ymm9[6,7,8,9],ymm14[10],ymm9[11,12],ymm14[13],ymm9[14,15]
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm4
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,1,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm30[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm23[0,2,2,3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm15[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
+; AVX512-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512-NEXT: vpermd %zmm2, %zmm1, %zmm27
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm1
+; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm20
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm5
+; AVX512-NEXT: vpshufb %xmm1, %xmm6, %xmm14
+; AVX512-NEXT: vmovdqa %xmm1, %xmm4
+; AVX512-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512-NEXT: vmovdqa (%rsi), %xmm2
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-NEXT: vprold $16, %xmm2, %xmm2
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm24
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm2 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm11 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6,7,8],ymm2[9],ymm1[10,11],ymm2[12],ymm1[13,14,15]
+; AVX512-NEXT: vmovdqa64 %ymm16, %ymm4
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
+; AVX512-NEXT: vprold $16, %ymm16, %ymm1
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm17[1,2,2,3,5,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm7 = ymm10[0,1],ymm1[2],ymm10[3,4],ymm1[5],ymm10[6,7,8,9],ymm1[10],ymm10[11,12],ymm1[13],ymm10[14,15]
+; AVX512-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,1,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm5[0,0,1,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm29[0,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm30[0,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,1,3,3]
+; AVX512-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm6 = mem[2,1,3,2]
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm5 = mem[2,2,2,3]
+; AVX512-NEXT: vpermq $232, (%rsp), %ymm1 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm1 = mem[0,2,2,3]
+; AVX512-NEXT: vmovdqa64 %xmm18, %xmm0
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
+; AVX512-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm0 = mem[2,1,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm14[0,0,1,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm19[2,2,2,3]
+; AVX512-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm15 = mem[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,1,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm28
+; AVX512-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm29
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %zmm28, %zmm30, %zmm29
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,1,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm1, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm30, %zmm9
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm0 # 32-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm0 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm1 = mem[2,3,3,3,6,7,7,7]
+; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm3 = mem[0,0,2,1]
+; AVX512-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm10 = mem[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,3]
+; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm12 = mem[0,0,1,1]
+; AVX512-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm14 = mem[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm17 = mem[2,2,2,3]
+; AVX512-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm18 = mem[2,1,3,2]
+; AVX512-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm19 = mem[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,0,1,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm24[0,0,2,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm23[2,1,3,2]
+; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm22[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm19[2,1,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm18[0,2,2,3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm5[2,1,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm4 = ymm8[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm12 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3],ymm12[4],ymm5[5,6,7,8],ymm12[9],ymm5[10,11],ymm12[12],ymm5[13,14,15]
-; AVX512-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm12
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm11, %zmm7
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $184, %zmm12, %zmm11, %zmm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm30, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm31, %zmm1
-; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm11, %zmm1
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm9 # 32-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm11, %zmm9
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm10[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm10 = mem[2,1,3,2]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm12 = mem[2,2,2,3]
-; AVX512-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm15 = mem[2,3,3,3,6,7,7,7]
-; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm17 = mem[0,0,2,1]
-; AVX512-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm8 = mem[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm18 = ymm26[0,0,1,1]
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm13
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm19 = ymm24[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm22[2,1,3,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm21[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm20[0,0,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm16[0,0,2,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,1,3,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,2,2,3]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm0
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm9
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm9
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm3
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm15[2,1,3,2]
-; AVX512-NEXT: vpbroadcastd 32(%rax), %ymm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm17, %zmm1
-; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm18, %zmm8
-; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm11, %zmm8
-; AVX512-NEXT: vpbroadcastd 36(%rax), %ymm1
-; AVX512-NEXT: vpbroadcastd 40(%rax), %ymm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm19, %zmm8, %zmm8
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm28, %zmm9, %zmm9
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,2,3],zmm29[0,1,2,3]
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm8
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512-NEXT: vpbroadcastd (%rax), %ymm9
-; AVX512-NEXT: vpbroadcastd 4(%rax), %ymm10
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm9
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm2
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm5
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm27
+; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm15[0,0,0,1]
; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512-NEXT: vpermd (%rax), %zmm2, %zmm2
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,2]
+; AVX512-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm4
+; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm4
+; AVX512-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512-NEXT: vpbroadcastd 40(%rax), %ymm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm2
; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm18, %zmm4, %zmm4
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm19[0,1,2,3]
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm20
+; AVX512-NEXT: vinserti64x4 $1, %ymm28, %zmm21, %zmm3
+; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm3
+; AVX512-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm4
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm5
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm3
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-NEXT: vpermd (%rax), %zmm6, %zmm6
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm6, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm6, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm20, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)
; AVX512-NEXT: vmovdqa64 %zmm0, 64(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm3, 384(%rax)
-; AVX512-NEXT: addq $632, %rsp # imm = 0x278
+; AVX512-NEXT: vmovdqa64 %zmm27, 384(%rax)
+; AVX512-NEXT: addq $680, %rsp # imm = 0x2A8
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i16_stride7_vf32:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $248, %rsp
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512-FCP-NEXT: subq $264, %rsp # imm = 0x108
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm2
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm2
-; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm10
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm10
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm3
; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm18
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm4
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm2
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
; AVX512-FCP-NEXT: vmovdqa %ymm4, %ymm9
-; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm6
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm17
; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm13
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm6
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm14
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm7
-; AVX512-FCP-NEXT: vporq %ymm6, %ymm7, %ymm25
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm6
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm8, %ymm5
-; AVX512-FCP-NEXT: vpor %ymm6, %ymm5, %ymm3
-; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
+; AVX512-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm15
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm4
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm11
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm11, %ymm8
+; AVX512-FCP-NEXT: vpor %ymm4, %ymm8, %ymm3
+; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm12
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm12, %ymm5
+; AVX512-FCP-NEXT: vpor %ymm1, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm14, %ymm1
; 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(%r9), %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm22
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm11, %ymm30
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm24
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm18[3,3,3,3,7,7,7,7]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,2,2,3,8,0,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm23
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
+; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
+; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm27
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
+; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm20
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7,8,9],ymm4[10],ymm0[11,12],ymm4[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [2,2,2,3,8,8,8,9]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm7
-; AVX512-FCP-NEXT: vpermi2q %zmm7, %zmm4, %zmm22
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vmovdqa %xmm3, %xmm4
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm21
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,8,8,8,9]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm6
+; AVX512-FCP-NEXT: vpermi2q %zmm6, %zmm0, %zmm23
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
+; AVX512-FCP-NEXT: vprold $16, %xmm10, %xmm6
; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm2
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm26
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm2
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm26
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm1
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm29
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm4
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,1,0,8,8,9,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm3, %zmm27
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm6
; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm9
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,0,1,8,9,9,11]
-; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm27
-; AVX512-FCP-NEXT: vprold $16, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
-; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm3 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm13[3],ymm4[4,5],ymm13[6],ymm4[7,8,9,10],ymm13[11],ymm4[12,13],ymm13[14],ymm4[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm28
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm2[4],xmm9[5],xmm2[5],xmm9[6],xmm2[6],xmm9[7],xmm2[7]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm0, %zmm4
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,0,1,1,12,13,0,15]
-; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm2, %zmm25
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm7
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,8,9,9,11]
+; AVX512-FCP-NEXT: vpermi2q %zmm7, %zmm3, %zmm28
+; AVX512-FCP-NEXT: vprold $16, %ymm15, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm5
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm5, %zmm31
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm2
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm9
+; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm10
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm9, %ymm4
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm31
-; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm7
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm15, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm21, %zmm29
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm8, %ymm1
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm21 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
-; AVX512-FCP-NEXT: vpermd %ymm8, %ymm21, %ymm2
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
-; AVX512-FCP-NEXT: vmovdqa %ymm7, %ymm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm13
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm13[0,1],ymm12[2],ymm13[3,4],ymm12[5],ymm13[6,7,8,9],ymm12[10],ymm13[11,12],ymm12[13],ymm13[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm30, %zmm12
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm6[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm13[2],ymm3[3,4],ymm13[5],ymm3[6,7,8,9],ymm13[10],ymm3[11,12],ymm13[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vprold $16, %ymm15, %ymm13
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm13[2],ymm6[3,4],ymm13[5],ymm6[6,7,8,9],ymm13[10],ymm6[11,12],ymm13[13],ymm6[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm13, %zmm6
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm15 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm10, %xmm10
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm10[1],xmm5[2,3],xmm10[4],xmm5[5,6],xmm10[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,0,1,1,8,8,10,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm2, %zmm5
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm3 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm3 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm1[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vmovdqa %xmm11, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm3, %xmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm10
-; AVX512-FCP-NEXT: vprold $16, %xmm14, %xmm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm3[2],xmm11[3,4],xmm3[5],xmm11[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm11
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm2, %zmm11
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm2
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm3[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm14[0],xmm2[1],xmm14[2,3],xmm2[4],xmm14[5,6],xmm2[7]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm24, %zmm2, %zmm12
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm6
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm12, %ymm3
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm12[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7,8,9],ymm4[10],ymm7[11,12],ymm4[13],ymm7[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm13, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm30
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7,8,9,10],ymm4[11],ymm0[12,13],ymm4[14],ymm0[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm22, %zmm9
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
+; AVX512-FCP-NEXT: vpermd %ymm2, %ymm22, %ymm4
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm12[0],ymm4[1],ymm12[2,3],ymm4[4],ymm12[5,6,7,8],ymm4[9],ymm12[10,11],ymm4[12],ymm12[13,14,15]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm12
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm8[2],ymm12[3,4],ymm8[5],ymm12[6,7,8,9],ymm8[10],ymm12[11,12],ymm8[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm20, %zmm12
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm8[2],ymm4[3,4],ymm8[5],ymm4[6,7,8,9],ymm8[10],ymm4[11,12],ymm8[13],ymm4[14,15]
+; AVX512-FCP-NEXT: vprold $16, %ymm13, %ymm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm8[2],ymm15[3,4],ymm8[5],ymm15[6,7,8,9],ymm8[10],ymm15[11,12],ymm8[13],ymm15[14,15]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm14[0],xmm8[1],xmm14[2,3],xmm8[4],xmm14[5,6],xmm8[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm14, %zmm8
+; AVX512-FCP-NEXT: vprold $16, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm5[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm7, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm14, %zmm3
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm15
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm4
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm13[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm6
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm24, %zmm16, %zmm12
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm15
+; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm5
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [4,5,2,2,6,6,6,6,30,31,27,27,31,31,30,31]
-; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm8, %zmm7
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
-; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm6
-; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm6
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm26, %zmm8, %zmm5
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm0
-; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm10
+; AVX512-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm7
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm26, %zmm2, %zmm8
+; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm12
+; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm12
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm12
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm27
+; AVX512-FCP-NEXT: vmovdqu (%rsp), %ymm4 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm24[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2,3],ymm3[4],ymm8[5,6,7,8],ymm3[9],ymm8[10,11],ymm3[12],ymm8[13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm8
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm18[0,0,1,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm17[2,2,2,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm15[0,1],ymm4[2],ymm15[3,4],ymm4[5],ymm15[6,7,8,9],ymm4[10],ymm15[11,12],ymm4[13],ymm15[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
+; AVX512-FCP-NEXT: vpermd %ymm10, %ymm3, %ymm3
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm14[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm19[0,0,1,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm18[2,2,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm17[2,1,3,2]
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm5
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm0[1],ymm14[2,3],ymm0[4],ymm14[5,6,7,8],ymm0[9],ymm14[10,11],ymm0[12],ymm14[13,14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm1[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7,8,9],ymm15[10],ymm14[11,12],ymm15[13],ymm14[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm14
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm14
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [6,0,0,0,7,0,0,7]
-; AVX512-FCP-NEXT: vpermd %ymm9, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm3
-; AVX512-FCP-NEXT: vpermd %zmm3, %zmm21, %zmm3
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm4 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm9 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm8, %zmm9
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm23, %zmm2, %zmm20
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm31
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm31
+; AVX512-FCP-NEXT: vpermd %zmm5, %zmm22, %zmm5
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm31, %zmm5
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm2, %zmm6
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm25, %zmm16, %zmm21
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm0
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm4
; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm16
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm2
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm2
+; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm3
+; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm3
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 320(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 256(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 64(%rax)
-; AVX512-FCP-NEXT: addq $248, %rsp
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512-FCP-NEXT: addq $264, %rsp # imm = 0x108
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i16_stride7_vf32:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $632, %rsp # imm = 0x278
+; AVX512DQ-NEXT: subq $680, %rsp # imm = 0x2A8
; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm1, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm29
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm27
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm14 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb %ymm14, %ymm8, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm9
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm1, %ymm0
-; AVX512DQ-NEXT: vmovdqa %ymm1, %ymm15
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm14 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm9, %ymm0
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm11
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm15 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm11, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm10
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm10, %ymm2
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm11
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm11, %ymm3
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm1, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm4
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm17
; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm3
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm7 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm7
-; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,6]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm10
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3]
+; AVX512DQ-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,6]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
-; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm3, %zmm4
+; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm4
; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm6
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm6, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm3, %ymm4
-; AVX512DQ-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm10
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm10, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm2, %ymm4
+; AVX512DQ-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm12
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm7
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm12, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm7, %ymm4
-; AVX512DQ-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm13, %ymm12, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm14, %ymm15, %ymm4
+; AVX512DQ-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm13
; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm14
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm14, %ymm0
-; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm2 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm8, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,17,17,0,0,0,0,0,1,2,0,0,3]
-; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm1, %zmm4
-; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,5,7,6]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm9 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
-; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vprold $16, %ymm10, %ymm4
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm9[0,1,2],ymm5[3],ymm9[4,5],ymm5[6],ymm9[7,8,9,10],ymm5[11],ymm9[12,13],ymm5[14],ymm9[15]
-; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm5, %zmm27
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm1 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,2]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3],xmm4[4],xmm5[5,6],xmm4[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm26 = [0,16,0,0,17,17,0,0,0,0,0,1,2,0,0,3]
+; AVX512DQ-NEXT: vpermi2d %zmm5, %zmm4, %zmm26
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm4
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm6[0,1,2,3,4,5,7,6]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
+; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm6, %zmm25
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm31
-; AVX512DQ-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vpbroadcastd 8(%rax), %ymm0
; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm5, %ymm4, %ymm4
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm28
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; AVX512DQ-NEXT: vmovdqa (%rax), %ymm3
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm7 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm3, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %ymm7, %ymm21
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm11 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512DQ-NEXT: vprold $16, %ymm13, %ymm0
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm13[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm28
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm29
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm30
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512DQ-NEXT: vprold $16, %ymm13, %ymm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm14[1,2,2,3,5,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm13[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm27, %ymm3
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm4
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm29[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512DQ-NEXT: vmovdqa %ymm15, %ymm8
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512DQ-NEXT: vprold $16, %xmm9, %xmm1
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm15 = xmm0[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0,1],xmm1[2],xmm15[3,4],xmm1[5],xmm15[6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512DQ-NEXT: vprold $16, %xmm5, %xmm6
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2],xmm7[3,4],xmm6[5],xmm7[6,7]
+; AVX512DQ-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm4, %xmm1
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[1,1,2,2]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm7[0],xmm1[1],xmm7[2,3],xmm1[4],xmm7[5,6],xmm1[7]
; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
-; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm9
-; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm15
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[1,1,2,2]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0],xmm0[1],xmm10[2,3],xmm0[4],xmm10[5,6],xmm0[7]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm26
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm9, %xmm25
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1],ymm3[2],ymm6[3,4],ymm3[5],ymm6[6,7,8,9],ymm3[10],ymm6[11,12],ymm3[13],ymm6[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm24
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm12[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1,2],ymm3[3],ymm6[4,5],ymm3[6],ymm6[7,8,9,10],ymm3[11],ymm6[12,13],ymm3[14],ymm6[15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm21
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm6
-; AVX512DQ-NEXT: vpermd %zmm6, %zmm3, %zmm3
-; AVX512DQ-NEXT: vmovdqa64 %ymm28, %ymm7
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm6, %ymm10
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,2,3]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm6
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm2, %xmm10
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm13
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512DQ-NEXT: vprold $16, %xmm14, %xmm14
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm13[0,1],xmm14[2],xmm13[3,4],xmm14[5],xmm13[6,7]
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm16
-; AVX512DQ-NEXT: vmovdqa %ymm4, %ymm2
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vmovdqa64 %ymm29, %ymm12
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm29[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm14 = ymm9[0,1],ymm14[2],ymm9[3,4],ymm14[5],ymm9[6,7,8,9],ymm14[10],ymm9[11,12],ymm14[13],ymm9[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm4
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,1,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm30[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm29 = ymm23[0,2,2,3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm15[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512DQ-NEXT: vpermd %zmm2, %zmm1, %zmm27
+; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm20
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm5
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm6, %xmm14
+; AVX512DQ-NEXT: vmovdqa %xmm1, %xmm4
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm2
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-NEXT: vprold $16, %xmm2, %xmm2
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm24
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm2 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm11 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6,7,8],ymm2[9],ymm1[10,11],ymm2[12],ymm1[13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm4
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
+; AVX512DQ-NEXT: vprold $16, %ymm16, %ymm1
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm17[1,2,2,3,5,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm7 = ymm10[0,1],ymm1[2],ymm10[3,4],ymm1[5],ymm10[6,7,8,9],ymm1[10],ymm10[11,12],ymm1[13],ymm10[14,15]
+; AVX512DQ-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,1,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm5[0,0,1,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm29 = ymm29[0,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm30[0,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,1,3,3]
+; AVX512DQ-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm6 = mem[2,1,3,2]
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm5 = mem[2,2,2,3]
+; AVX512DQ-NEXT: vpermq $232, (%rsp), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm1 = mem[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm0
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm12 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
+; AVX512DQ-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm0 = mem[2,1,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm14[0,0,1,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm19[2,2,2,3]
+; AVX512DQ-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm15 = mem[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,1,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm28
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm29
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm28, %zmm30, %zmm29
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,1,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm1, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm30, %zmm9
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512DQ-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm1 = mem[2,3,3,3,6,7,7,7]
+; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm3 = mem[0,0,2,1]
+; AVX512DQ-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm10 = mem[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,3]
+; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm12 = mem[0,0,1,1]
+; AVX512DQ-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm14 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm17 = mem[2,2,2,3]
+; AVX512DQ-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm18 = mem[2,1,3,2]
+; AVX512DQ-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm19 = mem[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm21 = ymm21[0,0,1,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm24[0,0,2,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm23[2,1,3,2]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm22[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm19[2,1,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm18[0,2,2,3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm5[2,1,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm4 = ymm8[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm12 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm12[1],ymm5[2,3],ymm12[4],ymm5[5,6,7,8],ymm12[9],ymm5[10,11],ymm12[12],ymm5[13,14,15]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm12
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm11, %zmm7
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %zmm12, %zmm11, %zmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm30, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm31, %zmm1
-; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm11, %zmm1
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm9 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm11, %zmm9
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm10[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm10 = mem[2,1,3,2]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm12 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm15 = mem[2,3,3,3,6,7,7,7]
-; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm17 = mem[0,0,2,1]
-; AVX512DQ-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm8 = mem[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm18 = ymm26[0,0,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm13
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm19 = ymm24[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm22[2,1,3,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm29 = ymm21[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm20[0,0,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm16[0,0,2,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,1,3,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,2,2,3]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm0
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm9
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm9
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm3
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm15[2,1,3,2]
-; AVX512DQ-NEXT: vpbroadcastd 32(%rax), %ymm9
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm17, %zmm1
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm18, %zmm8
-; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm11, %zmm8
-; AVX512DQ-NEXT: vpbroadcastd 36(%rax), %ymm1
-; AVX512DQ-NEXT: vpbroadcastd 40(%rax), %ymm9
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm19, %zmm8, %zmm8
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm28, %zmm9, %zmm9
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,2,3],zmm29[0,1,2,3]
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm8
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512DQ-NEXT: vpbroadcastd (%rax), %ymm9
-; AVX512DQ-NEXT: vpbroadcastd 4(%rax), %ymm10
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm9
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm5
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm27
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm15[0,0,0,1]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512DQ-NEXT: vpermd (%rax), %zmm2, %zmm2
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,2]
+; AVX512DQ-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm4
+; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm4
+; AVX512DQ-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512DQ-NEXT: vpbroadcastd 40(%rax), %ymm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm2
; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm18, %zmm4, %zmm4
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm19[0,1,2,3]
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm20
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm28, %zmm21, %zmm3
+; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm3
+; AVX512DQ-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512DQ-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm4
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm5
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm3
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-NEXT: vpermd (%rax), %zmm6, %zmm6
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm6, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm6, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm20, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm0, 64(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm3, 384(%rax)
-; AVX512DQ-NEXT: addq $632, %rsp # imm = 0x278
+; AVX512DQ-NEXT: vmovdqa64 %zmm27, 384(%rax)
+; AVX512DQ-NEXT: addq $680, %rsp # imm = 0x2A8
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i16_stride7_vf32:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $248, %rsp
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-FCP-NEXT: subq $264, %rsp # imm = 0x108
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm10
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm10
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm18
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm4, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT: vmovdqa %ymm4, %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm6
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm17
; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm13
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm7
-; AVX512DQ-FCP-NEXT: vporq %ymm6, %ymm7, %ymm25
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm11, %ymm8
+; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm8, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm12
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm12, %ymm5
+; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm14, %ymm1
; 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(%r9), %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm22
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm11, %ymm30
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm24
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm18[3,3,3,3,7,7,7,7]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,2,2,3,8,0,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm23
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm27
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm20
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7,8,9],ymm4[10],ymm0[11,12],ymm4[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm7
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm7, %zmm4, %zmm22
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, %xmm4
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm21
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm6
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm6, %zmm0, %zmm23
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm10, %xmm6
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm2
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm26
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm29
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,1,0,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm3, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm6
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm9
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm27
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} ymm3 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm13[3],ymm4[4,5],ymm13[6],ymm4[7,8,9,10],ymm13[11],ymm4[12,13],ymm13[14],ymm4[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm28
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm2[4],xmm9[5],xmm2[5],xmm9[6],xmm2[6],xmm9[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm0, %zmm4
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,0,1,1,12,13,0,15]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm2, %zmm25
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm7
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm7, %zmm3, %zmm28
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm15, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm5
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm5, %zmm31
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm9
+; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm10
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm9, %ymm4
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm31
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm7
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm15, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm21, %zmm29
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm8, %ymm1
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm21 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
-; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm21, %ymm2
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm13[0,1],ymm12[2],ymm13[3,4],ymm12[5],ymm13[6,7,8,9],ymm12[10],ymm13[11,12],ymm12[13],ymm13[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm30, %zmm12
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm6[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm13[2],ymm3[3,4],ymm13[5],ymm3[6,7,8,9],ymm13[10],ymm3[11,12],ymm13[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm15, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm13[2],ymm6[3,4],ymm13[5],ymm6[6,7,8,9],ymm13[10],ymm6[11,12],ymm13[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm13, %zmm6
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm15 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm10[1],xmm5[2,3],xmm10[4],xmm5[5,6],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,0,1,1,8,8,10,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm2, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm3 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm3 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm1[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm10
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm14, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm3[2],xmm11[3,4],xmm3[5],xmm11[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm11
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm2, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm3[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm14[0],xmm2[1],xmm14[2,3],xmm2[4],xmm14[5,6],xmm2[7]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm24, %zmm2, %zmm12
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm12, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm12[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7,8,9],ymm4[10],ymm7[11,12],ymm4[13],ymm7[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm13, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm10, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm30
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm6[0,1],ymm0[2],ymm6[3,4],ymm0[5],ymm6[6,7,8,9],ymm0[10],ymm6[11,12],ymm0[13],ymm6[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm4[3],ymm0[4,5],ymm4[6],ymm0[7,8,9,10],ymm4[11],ymm0[12,13],ymm4[14],ymm0[15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm22, %zmm9
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,1,4,5,4,5,5,7,12,13,10,10,14,14,14,14]
+; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm22, %ymm4
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm12[0],ymm4[1],ymm12[2,3],ymm4[4],ymm12[5,6,7,8],ymm4[9],ymm12[10,11],ymm4[12],ymm12[13,14,15]
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm8, %ymm12
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm8[2],ymm12[3,4],ymm8[5],ymm12[6,7,8,9],ymm8[10],ymm12[11,12],ymm8[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm20, %zmm12
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm8[2],ymm4[3,4],ymm8[5],ymm4[6,7,8,9],ymm8[10],ymm4[11,12],ymm8[13],ymm4[14,15]
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm13, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0,1],ymm8[2],ymm15[3,4],ymm8[5],ymm15[6,7,8,9],ymm8[10],ymm15[11,12],ymm8[13],ymm15[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm14[0],xmm8[1],xmm14[2,3],xmm8[4],xmm14[5,6],xmm8[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm14, %zmm8
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm5[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm7, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm14, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7,8,9],ymm6[10],ymm4[11,12],ymm6[13],ymm4[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm13[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm24, %zmm16, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm5
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [4,5,2,2,6,6,6,6,30,31,27,27,31,31,30,31]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm8, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
-; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm26, %zmm8, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm0
-; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm10
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm5, %zmm2, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm26, %zmm2, %zmm8
+; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm12
+; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqu (%rsp), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm24[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm8[0],ymm3[1],ymm8[2,3],ymm3[4],ymm8[5,6,7,8],ymm3[9],ymm8[10,11],ymm3[12],ymm8[13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm8
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm18[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm17[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm15[0,1],ymm4[2],ymm15[3,4],ymm4[5],ymm15[6,7,8,9],ymm4[10],ymm15[11,12],ymm4[13],ymm15[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm3, %ymm3
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm14[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm19[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm18[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm17[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm0[1],ymm14[2,3],ymm0[4],ymm14[5,6,7,8],ymm0[9],ymm14[10,11],ymm0[12],ymm14[13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm1[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7,8,9],ymm15[10],ymm14[11,12],ymm15[13],ymm14[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm14
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm14
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [6,0,0,0,7,0,0,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm9, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT: vpermd %zmm3, %zmm21, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm9 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm8, %zmm9
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm23, %zmm2, %zmm20
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm31
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm31
+; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm22, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm31, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm2, %zmm6
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm25, %zmm16, %zmm21
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm4
; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm16
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm2
+; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm3
+; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm3
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 320(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 256(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 64(%rax)
-; AVX512DQ-FCP-NEXT: addq $248, %rsp
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-FCP-NEXT: addq $264, %rsp # imm = 0x108
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -12619,2682 +12639,2656 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride7_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $2168, %rsp # imm = 0x878
-; AVX512-NEXT: vmovdqa 96(%rcx), %ymm2
-; AVX512-NEXT: vmovdqa 96(%rdx), %ymm9
+; AVX512-NEXT: subq $2808, %rsp # imm = 0xAF8
+; AVX512-NEXT: vmovdqa 96(%rcx), %ymm6
+; AVX512-NEXT: vmovdqa 96(%rdx), %ymm13
; AVX512-NEXT: vmovdqa 96(%rdi), %ymm7
; AVX512-NEXT: vmovdqa 96(%rsi), %ymm8
; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-NEXT: vpshufb %ymm11, %ymm9, %ymm2
-; AVX512-NEXT: vporq %ymm1, %ymm2, %ymm16
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm12, %ymm8, %ymm1
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-NEXT: vpshufb %ymm13, %ymm7, %ymm2
-; AVX512-NEXT: vporq %ymm1, %ymm2, %ymm17
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm0, %ymm6, %ymm2
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512-NEXT: vpshufb %ymm10, %ymm8, %ymm2
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512-NEXT: vpshufb %ymm11, %ymm7, %ymm3
+; AVX512-NEXT: vporq %ymm2, %ymm3, %ymm18
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
; AVX512-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa %ymm3, %ymm10
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm12, %ymm2, %ymm2
; AVX512-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-NEXT: vpshufb %ymm15, %ymm3, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm20
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 64(%rcx), %ymm2
-; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512-NEXT: vmovdqa 64(%rdx), %ymm6
-; AVX512-NEXT: vpshufb %ymm11, %ymm6, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 64(%rsi), %ymm5
-; AVX512-NEXT: vpshufb %ymm12, %ymm5, %ymm1
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512-NEXT: vpshufb %ymm9, %ymm3, %ymm3
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 64(%rcx), %ymm3
+; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512-NEXT: vmovdqa 64(%rdx), %ymm4
+; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm27
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 64(%rsi), %ymm3
+; AVX512-NEXT: vpshufb %ymm10, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm23
; AVX512-NEXT: vmovdqa 64(%rdi), %ymm4
-; AVX512-NEXT: vpshufb %ymm13, %ymm4, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm10, %ymm1, %ymm1
-; AVX512-NEXT: vmovdqa %ymm10, %ymm3
-; AVX512-NEXT: vmovdqa (%r8), %ymm2
+; AVX512-NEXT: vpshufb %ymm11, %ymm4, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm22
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm15, %ymm2, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm0, %ymm1, %ymm1
-; AVX512-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512-NEXT: vmovdqa (%r9), %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm12, %ymm2, %ymm2
+; AVX512-NEXT: vmovdqa (%r8), %ymm3
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm9, %ymm3, %ymm3
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm2
+; AVX512-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm1, %ymm3, %ymm3
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm11, %ymm2, %ymm2
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512-NEXT: vpshufb %ymm12, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512-NEXT: vpshufb %ymm13, %ymm10, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm10, %ymm22
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm0
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512-NEXT: vpor %ymm0, %ymm11, %ymm0
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm10, %ymm2, %ymm2
+; AVX512-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX512-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512-NEXT: vpshufb %ymm12, %ymm0, %ymm12
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512-NEXT: vpshufb %ymm13, %ymm11, %ymm13
-; AVX512-NEXT: vpor %ymm12, %ymm13, %ymm10
-; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512-NEXT: vpshufb %ymm15, %ymm13, %ymm12
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512-NEXT: vpshufb %ymm3, %ymm15, %ymm14
-; AVX512-NEXT: vpor %ymm14, %ymm12, %ymm10
-; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vprold $16, %ymm15, %ymm12
-; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm13[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0,1],ymm12[2],ymm14[3,4],ymm12[5],ymm14[6,7,8,9],ymm12[10],ymm14[11,12],ymm12[13],ymm14[14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm14 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512-NEXT: vpshufb %ymm10, %ymm5, %ymm0
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512-NEXT: vpshufb %ymm11, %ymm3, %ymm1
+; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512-NEXT: vpshufb %ymm9, %ymm1, %ymm9
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm0
+; AVX512-NEXT: vpshufb %ymm12, %ymm0, %ymm10
+; AVX512-NEXT: vpor %ymm10, %ymm9, %ymm9
+; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm10, %ymm6, %ymm9
+; AVX512-NEXT: vmovdqa64 %ymm10, %ymm31
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm9 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm14[0,1],ymm10[2],ymm14[3,4],ymm10[5],ymm14[6,7,8,9],ymm10[10],ymm14[11,12],ymm10[13],ymm14[14,15]
-; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
-; AVX512-NEXT: vpermi2q %zmm10, %zmm12, %zmm14
-; AVX512-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm14
-; AVX512-NEXT: vpshufb %ymm12, %ymm14, %ymm10
-; AVX512-NEXT: vmovdqa64 %ymm12, %ymm19
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm12[2],ymm10[3,4],ymm12[5],ymm10[6,7,8,9],ymm12[10],ymm10[11,12],ymm12[13],ymm10[14,15]
-; AVX512-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
-; AVX512-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
-; AVX512-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm9 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm9 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3],ymm10[4,5],ymm9[6],ymm10[7,8,9,10],ymm9[11],ymm10[12,13],ymm9[14],ymm10[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,3,3]
+; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 96(%r9), %ymm9
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm9[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
; AVX512-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm14[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,3,2]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm9[2,1,2,3,6,5,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm10[2,2,2,2]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm6 = ymm10[0,1],ymm6[2],ymm10[3,4],ymm6[5],ymm10[6,7,8,9],ymm6[10],ymm10[11,12],ymm6[13],ymm10[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[1,1,1,1,5,5,5,5]
; AVX512-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,2,1,4,4,6,5]
; AVX512-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4],ymm7[5],ymm8[6,7,8,9],ymm7[10],ymm8[11,12],ymm7[13],ymm8[14,15]
; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm16, %zmm8
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm17, %zmm7
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm7
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512-NEXT: vmovdqa 96(%r8), %ymm10
+; AVX512-NEXT: vprold $16, %ymm9, %ymm8
+; AVX512-NEXT: vpshufb %ymm12, %ymm9, %ymm9
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm17, %zmm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm7
+; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
+; AVX512-NEXT: vmovdqa 96(%r8), %ymm6
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[1,2,2,3,5,6,6,7]
; AVX512-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm8 = ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17,u,u,u,u],zero,zero
-; AVX512-NEXT: vpternlogq $248, %ymm26, %ymm7, %ymm8
-; AVX512-NEXT: vmovdqa 96(%r9), %ymm12
-; AVX512-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %ymm3, %ymm12, %ymm9
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $248, %ymm14, %ymm8, %ymm9
-; AVX512-NEXT: vextracti64x4 $1, %zmm7, %ymm7
-; AVX512-NEXT: vpshufd {{.*#+}} ymm8 = ymm10[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[16,17,u,u,u,u],zero,zero
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $248, %ymm11, %ymm7, %ymm6
+; AVX512-NEXT: vmovdqa64 %zmm11, %zmm12
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $248, %ymm11, %ymm6, %ymm9
+; AVX512-NEXT: vextracti64x4 $1, %zmm7, %ymm6
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm10[2,1,3,3]
; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512-NEXT: vpternlogq $184, %ymm7, %ymm10, %ymm8
-; AVX512-NEXT: vprold $16, %ymm12, %ymm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $184, %ymm8, %ymm14, %ymm7
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm9[0,1,2,3],zmm7[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpternlogq $184, %ymm6, %ymm10, %ymm7
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm8[2,2,2,2]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %ymm7, %ymm8, %ymm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,2,3],zmm6[4,5,6,7]
+; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
-; AVX512-NEXT: vmovdqa 96(%rax), %ymm7
-; AVX512-NEXT: vpermd %zmm7, %zmm18, %zmm9
-; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm10 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-NEXT: vpshufb %ymm10, %ymm7, %ymm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
-; AVX512-NEXT: vpandnq %ymm9, %ymm26, %ymm9
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512-NEXT: vpermd %zmm6, %zmm18, %zmm7
; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-NEXT: vpbroadcastd 72(%rax), %ymm7
-; AVX512-NEXT: vpandnq %ymm7, %ymm16, %ymm9
+; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-NEXT: vpshufb %ymm11, %ymm6, %ymm6
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,2,3]
+; AVX512-NEXT: vpandn %ymm7, %ymm12, %ymm7
+; AVX512-NEXT: vmovdqa64 %zmm12, %zmm19
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512-NEXT: vpbroadcastd 72(%rax), %ymm6
+; AVX512-NEXT: vpandnq %ymm6, %ymm28, %ymm6
; AVX512-NEXT: vmovdqa 64(%rax), %ymm7
-; AVX512-NEXT: vpshufb %ymm10, %ymm7, %ymm12
-; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
-; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpbroadcastd 8(%rax), %ymm9
-; AVX512-NEXT: vpandnq %ymm9, %ymm16, %ymm9
-; AVX512-NEXT: vmovdqa (%rax), %ymm8
-; AVX512-NEXT: vpshufb %ymm10, %ymm8, %ymm12
-; AVX512-NEXT: vmovdqa64 %ymm8, %ymm24
-; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
-; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0,1],ymm9[2],ymm12[3,4],ymm9[5],ymm12[6,7,8,9],ymm9[10],ymm12[11,12],ymm9[13],ymm12[14,15]
-; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm12[2],ymm9[3,4],ymm12[5],ymm9[6,7,8,9],ymm12[10],ymm9[11,12],ymm12[13],ymm9[14,15]
-; AVX512-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512-NEXT: vpshufb %ymm10, %ymm9, %ymm10
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm11, %ymm7, %ymm7
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 64(%r9), %xmm7
+; AVX512-NEXT: vmovdqa 64(%r8), %xmm8
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512-NEXT: vmovdqa64 %xmm8, %xmm17
+; AVX512-NEXT: vmovdqa64 %xmm7, %xmm20
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-NEXT: vpshufb %xmm7, %xmm6, %xmm6
+; AVX512-NEXT: vmovdqa %xmm7, %xmm15
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 64(%rcx), %xmm9
+; AVX512-NEXT: vmovdqa 64(%rdx), %xmm7
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 64(%rdi), %xmm14
+; AVX512-NEXT: vmovdqa 64(%rsi), %xmm10
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
+; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpbroadcastd 8(%rax), %ymm8
+; AVX512-NEXT: vpandnq %ymm8, %ymm28, %ymm8
+; AVX512-NEXT: vmovdqa (%rax), %ymm12
+; AVX512-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm11, %ymm12, %ymm13
+; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
+; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%r9), %xmm6
+; AVX512-NEXT: vmovdqa (%r8), %xmm12
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512-NEXT: vmovdqa64 %xmm12, %xmm29
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm24
+; AVX512-NEXT: vpshufb %xmm15, %xmm8, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm15, %xmm25
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
+; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%rcx), %xmm8
+; AVX512-NEXT: vmovdqa (%rdx), %xmm13
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,0,2,1]
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512-NEXT: vmovdqa64 %xmm12, %xmm21
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm16
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,0,1,3]
+; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,3,2]
+; AVX512-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm12 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512-NEXT: vpshufb %ymm11, %ymm15, %ymm11
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm15[0,1,1,3,4,5,5,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,2,2,3]
-; AVX512-NEXT: vpandnq %ymm12, %ymm26, %ymm12
-; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10
-; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
-; AVX512-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm13
-; AVX512-NEXT: vpshufb %ymm13, %ymm2, %ymm10
-; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0,1],ymm12[2],ymm10[3,4],ymm12[5],ymm10[6,7,8,9],ymm12[10],ymm10[11,12],ymm12[13],ymm10[14,15]
-; AVX512-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpandnq %ymm12, %ymm19, %ymm12
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
+; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm11 = ymm0[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1,2],ymm12[3],ymm11[4,5],ymm12[6],ymm11[7,8,9,10],ymm12[11],ymm11[12,13],ymm12[14],ymm11[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
+; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm31, %ymm6
+; AVX512-NEXT: vpshufb %ymm6, %ymm4, %ymm11
+; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm12[2],ymm11[3,4],ymm12[5],ymm11[6,7,8,9],ymm12[10],ymm11[11,12],ymm12[13],ymm11[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6,7,8],ymm4[9],ymm2[10,11],ymm4[12],ymm2[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm2 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6,7,8],ymm2[9],ymm1[10,11],ymm2[12],ymm1[13,14,15]
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm31
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm3 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,3]
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vprold $16, %ymm0, %ymm2
+; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,2]
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpermd %zmm9, %zmm18, %zmm0
+; AVX512-NEXT: vpermd %zmm15, %zmm18, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512-NEXT: vmovdqa64 %ymm23, %ymm12
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[1,1,1,1,5,5,5,5]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm22[1,1,1,1,5,5,5,5]
; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm12[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm22[2,2,2,2,6,6,6,6]
; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm2
+; AVX512-NEXT: vmovdqa64 %ymm26, %ymm2
; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm27[0,1,1,3,4,5,5,7]
; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512-NEXT: vpshufb %ymm13, %ymm2, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm23, %ymm3
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm31, %ymm11
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm27[2,2,2,2,6,6,6,6]
; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm8
-; AVX512-NEXT: vprold $16, %ymm21, %ymm0
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm20[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,2,3,3,10,9,11,10]
-; AVX512-NEXT: vpermt2q %zmm0, %zmm10, %zmm1
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm11 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512-NEXT: vpermd 64(%rax), %zmm11, %zmm0
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm26, %zmm0
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vprold $16, %ymm5, %ymm0
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,2,2,3,5,6,6,7]
+; AVX512-NEXT: vmovdqa %ymm2, %ymm6
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm27[3,3,3,3,7,7,7,7]
; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6,7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13,14,15]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm8[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm12 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 96(%rcx), %xmm0
+; AVX512-NEXT: vmovdqa 96(%rdx), %xmm1
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm3 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,3]
+; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 96(%rsi), %xmm2
+; AVX512-NEXT: vmovdqa 96(%rdi), %xmm3
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-NEXT: vmovdqa64 %xmm25, %xmm12
+; AVX512-NEXT: vpshufb %xmm12, %xmm4, %xmm4
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vprold $16, %xmm2, %xmm4
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
+; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512-NEXT: vmovdqa %xmm3, %xmm4
+; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT: vmovdqa 96(%r9), %xmm0
; AVX512-NEXT: vmovdqa 96(%r8), %xmm1
; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpermt2d %zmm4, %zmm12, %zmm3
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[2,3,3,3,6,7,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,3,2]
-; AVX512-NEXT: vpbroadcastd 96(%rax), %ymm5
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-NEXT: vpternlogd $184, %zmm3, %zmm5, %zmm4
-; AVX512-NEXT: vmovdqa64 %zmm5, %zmm15
-; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 96(%rsi), %xmm3
-; AVX512-NEXT: vmovdqa 96(%rdi), %xmm4
-; AVX512-NEXT: vprold $16, %xmm3, %xmm5
-; AVX512-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
-; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-NEXT: vmovdqa64 %xmm5, %xmm23
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512-NEXT: vmovdqa 96(%rdx), %xmm4
-; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm6 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-NEXT: vpshufb %xmm6, %xmm3, %xmm5
-; AVX512-NEXT: vmovdqa %xmm6, %xmm7
-; AVX512-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,6]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm29 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
-; AVX512-NEXT: vpermt2d %zmm2, %zmm29, %zmm1
-; AVX512-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512-NEXT: vpbroadcastd 104(%rax), %ymm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm14, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqa %xmm7, %xmm8
-; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
+; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512-NEXT: vpermt2d %zmm1, %zmm18, %zmm0
+; AVX512-NEXT: vpbroadcastd 100(%rax), %ymm1
+; AVX512-NEXT: vpbroadcastd 104(%rax), %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm31
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm31
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
+; AVX512-NEXT: vpshufb %xmm4, %xmm9, %xmm1
+; AVX512-NEXT: vmovdqa %xmm4, %xmm6
+; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[1,1,2,2]
; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [16,16,17,17,17,17,0,0,0,1,0,1,2,3,2,3]
-; AVX512-NEXT: vpermt2d %zmm2, %zmm3, %zmm1
-; AVX512-NEXT: vmovdqa 64(%rdi), %xmm2
-; AVX512-NEXT: vmovdqa 64(%rsi), %xmm4
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX512-NEXT: vpshufb %xmm0, %xmm5, %xmm5
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512-NEXT: vmovdqa64 %xmm6, %xmm25
-; AVX512-NEXT: vprold $16, %xmm4, %xmm4
-; AVX512-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm4[2],xmm2[3,4],xmm4[5],xmm2[6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 64(%r9), %xmm1
-; AVX512-NEXT: vmovdqa 64(%r8), %xmm2
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-NEXT: vmovdqa64 %xmm5, %xmm17
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,7,6]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
-; AVX512-NEXT: vpermt2d %zmm2, %zmm6, %zmm1
-; AVX512-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512-NEXT: vpbroadcastd 68(%rax), %ymm5
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm9, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512-NEXT: vmovdqa (%rdx), %xmm5
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqa %xmm8, %xmm14
-; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm5[0],xmm1[1],xmm5[2,3],xmm1[4],xmm5[5,6],xmm1[7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm7[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpermt2d %zmm5, %zmm3, %zmm1
-; AVX512-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
-; AVX512-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm19
-; AVX512-NEXT: vprold $16, %xmm5, %xmm5
-; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm5[2],xmm3[3,4],xmm5[5],xmm3[6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm2 = [16,16,17,17,17,17,0,0,0,1,0,1,2,3,2,3]
+; AVX512-NEXT: vpermt2d %zmm0, %zmm2, %zmm1
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
+; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
+; AVX512-NEXT: vprold $16, %xmm10, %xmm3
+; AVX512-NEXT: vpshufd {{.*#+}} xmm4 = xmm14[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm2
-; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-NEXT: vmovdqa (%r8), %xmm3
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm14
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm0, %zmm14
+; AVX512-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm20, %xmm3
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,7,6]
; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpermt2d %zmm3, %zmm6, %zmm1
-; AVX512-NEXT: vpbroadcastd (%rax), %ymm3
-; AVX512-NEXT: vpbroadcastd 4(%rax), %ymm4
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm2
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm9, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm28, %ymm7
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm22[1,1,1,1,5,5,5,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm22
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm7[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vmovdqa %ymm2, %ymm9
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2,3],ymm1[4],ymm4[5,6,7,8],ymm1[9],ymm4[10,11],ymm1[12],ymm4[13,14,15]
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm21
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
+; AVX512-NEXT: vpermt2d %zmm3, %zmm4, %zmm1
+; AVX512-NEXT: vpbroadcastd 64(%rax), %ymm3
+; AVX512-NEXT: vpbroadcastd 68(%rax), %ymm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm25
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm19, %zmm25
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3]
+; AVX512-NEXT: vpshufb %xmm6, %xmm8, %xmm3
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm13[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpermt2d %zmm1, %zmm2, %zmm3
+; AVX512-NEXT: vmovdqa64 %xmm16, %xmm5
+; AVX512-NEXT: vmovdqa64 %xmm21, %xmm2
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512-NEXT: vpshufb %xmm12, %xmm1, %xmm1
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
+; AVX512-NEXT: vprold $16, %xmm21, %xmm2
+; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm16[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm5[0,1],xmm2[2],xmm5[3,4],xmm2[5],xmm5[6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm5
+; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm0, %zmm5
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,7,6]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpermt2d %zmm1, %zmm4, %zmm0
+; AVX512-NEXT: vpbroadcastd (%rax), %ymm1
+; AVX512-NEXT: vpbroadcastd 4(%rax), %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm20
+; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm19, %zmm20
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-NEXT: vprold $16, %ymm2, %ymm1
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vmovdqa %ymm2, %ymm10
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
+; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm4 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm3[0,0,2,1,4,4,6,5]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm6[3],ymm4[4,5],ymm6[6],ymm4[7,8,9,10],ymm6[11],ymm4[12,13],ymm6[14],ymm4[15]
-; AVX512-NEXT: vpermt2q %zmm1, %zmm10, %zmm4
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm6[0,1],ymm1[2],ymm6[3,4],ymm1[5],ymm6[6,7,8,9],ymm1[10],ymm6[11,12],ymm1[13],ymm6[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512-NEXT: vpshufb %ymm13, %ymm8, %ymm1
-; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm6[2],ymm1[3,4],ymm6[5],ymm1[6,7,8,9],ymm6[10],ymm1[11,12],ymm6[13],ymm1[14,15]
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512-NEXT: vpermd (%rax), %zmm11, %zmm1
-; AVX512-NEXT: vpternlogd $184, %zmm4, %zmm26, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2,3],ymm4[4],ymm1[5,6,7,8],ymm4[9],ymm1[10,11],ymm4[12],ymm1[13,14,15]
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm7[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm4[3],ymm1[4,5],ymm4[6],ymm1[7,8,9,10],ymm4[11],ymm1[12,13],ymm4[14],ymm1[15]
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm4 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm6[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpermt2d %zmm11, %zmm12, %zmm1
-; AVX512-NEXT: vpshufd {{.*#+}} ymm11 = ymm24[2,3,3,3,6,7,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512-NEXT: vpbroadcastd 32(%rax), %ymm12
-; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm26
-; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm15, %zmm26
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512-NEXT: vprold $16, %xmm12, %xmm15
-; AVX512-NEXT: vpshufd {{.*#+}} xmm13 = xmm1[1,1,2,3]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm2 = xmm13[0,1],xmm15[2],xmm13[3,4],xmm15[5],xmm13[6,7]
-; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm12[0],xmm1[1],xmm12[1],xmm1[2],xmm12[2],xmm1[3],xmm12[3]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
-; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
-; AVX512-NEXT: vmovdqa64 %xmm17, %xmm1
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512-NEXT: vpshufb %xmm0, %xmm5, %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm23, %xmm1
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %xmm0, %xmm13, %xmm1
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm24 = mem[2,2,2,3]
-; AVX512-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm23 = mem[0,2,2,3]
-; AVX512-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm7 = mem[0,2,2,3]
-; AVX512-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm17 = mem[2,1,3,3]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm6[0,1,2,3,4,5,7,6]
-; AVX512-NEXT: vpermt2d %zmm0, %zmm29, %zmm4
-; AVX512-NEXT: vpbroadcastd 36(%rax), %ymm0
-; AVX512-NEXT: vpbroadcastd 40(%rax), %ymm6
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm28
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm28
-; AVX512-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512-NEXT: vpshufb %xmm14, %xmm6, %xmm4
-; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512-NEXT: vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
-; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm12[0],xmm4[1],xmm12[2,3],xmm4[4],xmm12[5,6],xmm4[7]
-; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
-; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm0[3,3,3,3]
-; AVX512-NEXT: vpshufd $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm0 = mem[1,2,2,3,5,6,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm1[2,1,2,3,6,5,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,1]
-; AVX512-NEXT: vmovdqa64 %xmm27, %xmm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm1[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm1[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm15 = xmm1[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
-; AVX512-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm1[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,3]
-; AVX512-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm4 = mem[2,1,3,2]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm5 = mem[2,2,2,3]
-; AVX512-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm6 = mem[2,2,3,3]
-; AVX512-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm1 = mem[2,2,2,3]
-; AVX512-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm1 = mem[0,2,2,3]
-; AVX512-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,2,2,3]
-; AVX512-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm27 = mem[2,1,3,3]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm20 = mem[2,2,2,3]
-; AVX512-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm19 = ymm30[2,1,3,2]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm1 = mem[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm25 = ymm22[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm21[0,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,1,3,2]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm18[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm23
-; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm7, %zmm24
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $184, %zmm23, %zmm30, %zmm24
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[0,1,1,3,4,5,5,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-NEXT: vpternlogq $184, %ymm24, %ymm7, %ymm0
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-NEXT: vpternlogq $184, %ymm0, %ymm18, %ymm14
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm2 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm3 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $226, %zmm10, %zmm23, %zmm8
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm17 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm17, %zmm8
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm2 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm10 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm23, %zmm10
-; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm17, %zmm10
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %ymm11, %ymm2, %ymm0
+; AVX512-NEXT: vmovdqa %ymm2, %ymm9
+; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vmovdqa %ymm3, %ymm11
+; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[0,0,2,1,4,4,6,5]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0,1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8,9,10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
+; AVX512-NEXT: vprold $16, %ymm7, %ymm0
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[1,2,2,3,5,6,6,7]
+; AVX512-NEXT: vmovdqa %ymm4, %ymm8
+; AVX512-NEXT: vpblendw {{.*#+}} ymm4 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7,8,9],ymm0[10],ymm2[11,12],ymm0[13],ymm2[14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm2 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6,7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm10[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm6[3],ymm0[4,5],ymm6[6],ymm0[7,8,9,10],ymm6[11],ymm0[12,13],ymm6[14],ymm0[15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm6 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
+; AVX512-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpermq {{.*#+}} ymm22 = ymm1[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm26 = ymm3[2,2,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm24 = ymm4[2,1,3,2]
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512-NEXT: vprold $16, %xmm10, %xmm1
+; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[1,1,2,3]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2],xmm3[3,4],xmm1[5],xmm3[6,7]
+; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpermq {{.*#+}} ymm21 = ymm2[0,2,2,3]
+; AVX512-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm2
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX512-NEXT: vpshufb %xmm12, %xmm2, %xmm3
+; AVX512-NEXT: vpshufb %xmm12, %xmm1, %xmm1
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,4,5,7,6]
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm27
+; AVX512-NEXT: vpermt2d %zmm2, %zmm18, %zmm1
+; AVX512-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm13
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm13
+; AVX512-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm12 = xmm4[0,0,1,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm0[2,1,3,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm3[0,0,1,1]
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm2
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm21
-; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21
-; AVX512-NEXT: vextracti64x4 $1, %zmm24, %ymm2
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm2
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $184, %ymm2, %ymm3, %ymm29
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm22 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %zmm3, %zmm4, %zmm1
+; AVX512-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm3 = mem[2,1,3,2]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm29 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %ymm1, %ymm29, %ymm3
+; AVX512-NEXT: vpternlogq $184, %ymm3, %ymm28, %ymm30
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm0[0,1,2,3],zmm6[0,1,2,3]
-; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm24 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm2
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm14[0,1,2,3],zmm2[4,5,6,7]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm0, %zmm29
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm7 # 64-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm18 # 64-byte Folded Reload
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm17 # 64-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm20, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm19, %zmm3
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm25, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm16, %zmm5
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm5
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, %zmm3, %zmm2, %zmm20
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, %zmm5, %zmm2, %zmm25
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm3 # 64-byte Folded Reload
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm27, %zmm31, %zmm3
-; AVX512-NEXT: vpternlogq $184, %zmm2, %zmm30, %zmm3
-; AVX512-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm4 = mem[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,1,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm4 = mem[2,1,3,3]
-; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm5 = mem[0,0,1,1]
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm5 = mem[0,0,2,1]
-; AVX512-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm6 = mem[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
-; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm19 = mem[0,0,1,1]
-; AVX512-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm9 = mem[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
-; AVX512-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm13 = mem[2,1,3,3]
-; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm14 = mem[0,0,1,1]
-; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm15 = mem[0,0,2,1]
-; AVX512-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm0 = mem[2,1,2,3,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,3]
-; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm16 = mem[0,0,1,1]
-; AVX512-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm1 = mem[0,2,3,3,4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm30, %zmm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm12[0,1,1,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm11, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm11
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm30, %zmm11
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, %zmm4, %zmm2, %zmm12
-; AVX512-NEXT: vpternlogq $184, %zmm11, %zmm2, %zmm26
-; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm19, %zmm4
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm23, %zmm4
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm16, %zmm1
-; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm23, %zmm1
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm23 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm18 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $226, %zmm23, %zmm28, %zmm18
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm18
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm23 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm23
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm23
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm17 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm17
+; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm19, %ymm0
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-NEXT: vpternlogq $184, %ymm0, %ymm1, %ymm2
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
+; AVX512-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm19 # 64-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[0,1,2,3]
+; AVX512-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm1
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm30[0,1,2,3],zmm1[4,5,6,7]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm30
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm11
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %zmm14, %zmm1, %zmm25
+; AVX512-NEXT: vpternlogq $184, %zmm5, %zmm1, %zmm20
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm5, %zmm2
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm22, %zmm3, %zmm22
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm5, %zmm22
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-NEXT: vpermd 64(%rax), %zmm14, %zmm5
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm3, %zmm5
+; AVX512-NEXT: vinserti64x4 $1, %ymm24, %zmm26, %zmm1
+; AVX512-NEXT: vpermd (%rax), %zmm14, %zmm14
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm3, %zmm14
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm5
+; AVX512-NEXT: vpternlogq $184, %zmm22, %zmm1, %zmm14
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vpternlogq $184, %zmm4, %zmm0, %zmm2
-; AVX512-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm28
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm24
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm7
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $184, %zmm1, %zmm4, %zmm2
+; AVX512-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm1 = mem[0,1,1,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm22 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm22
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm12[0,1,1,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm21, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm15, %zmm8
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm8
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-NEXT: vpshufb {{.*#+}} xmm4 = xmm7[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[1,1,2,2]
+; AVX512-NEXT: vpblendw {{.*#+}} xmm4 = xmm9[0],xmm4[1],xmm9[2,3],xmm4[4],xmm9[5,6],xmm4[7]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm7 = mem[2,2,2,3]
+; AVX512-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm9 = mem[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[0,1,1,3]
+; AVX512-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm10 = mem[2,3,3,3,6,7,7,7]
+; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm16 = mem[0,0,2,1]
+; AVX512-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm12 = mem[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,3]
+; AVX512-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm21 = mem[0,0,1,1]
+; AVX512-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm15 = mem[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm24 = mem[2,2,2,3]
+; AVX512-NEXT: vmovdqa64 %xmm27, %xmm3
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,1,3]
+; AVX512-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm26 = mem[2,3,3,3,6,7,7,7]
+; AVX512-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm27 = mem[0,0,2,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,1,2,3,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,4]
+; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm10[2,1,3,2]
+; AVX512-NEXT: vpbroadcastd 96(%rax), %ymm10
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm9
+; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm24, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm26[2,1,3,2]
+; AVX512-NEXT: vpbroadcastd 32(%rax), %ymm10
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm7
+; AVX512-NEXT: vpternlogd $184, %zmm3, %zmm29, %zmm7
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512-NEXT: vpternlogq $184, %zmm22, %zmm3, %zmm9
+; AVX512-NEXT: vpternlogq $184, %zmm8, %zmm3, %zmm7
+; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm16, %zmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm8
+; AVX512-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm8
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm27, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm3
+; AVX512-NEXT: vpternlogq $226, %zmm1, %zmm28, %zmm3
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512-NEXT: vpternlogq $184, %zmm8, %zmm1, %zmm31
+; AVX512-NEXT: vpternlogq $184, %zmm3, %zmm1, %zmm13
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm0
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm11
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm24, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm28, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm26, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm25, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm17, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm18, 448(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm2, 704(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm12, 640(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm20, 576(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm8, 512(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm22, 768(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm29, 832(%rax)
-; AVX512-NEXT: addq $2168, %rsp # imm = 0x878
+; AVX512-NEXT: vmovdqa64 %zmm0, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm13, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm14, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm23, 64(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm20, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm25, 448(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm31, 704(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm9, 640(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm5, 576(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm18, 512(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm11, 384(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm19, 768(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm30, 832(%rax)
+; AVX512-NEXT: addq $2808, %rsp # imm = 0xAF8
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i16_stride7_vf64:
; AVX512-FCP: # %bb.0:
; AVX512-FCP-NEXT: subq $1496, %rsp # imm = 0x5D8
-; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm0
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %ymm9
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm3
-; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm10
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm16
-; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %ymm6
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm3
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %ymm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm2
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
+; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %ymm4
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm5
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm3
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm2
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm7
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm12
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm28
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm4
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm4, %ymm18
; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm14
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm31
; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm12
-; AVX512-FCP-NEXT: vpor %ymm8, %ymm12, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm15
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm14
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm15
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
+; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm13, %ymm19
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm15
+; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa %ymm10, %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm14
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm12
+; AVX512-FCP-NEXT: vpor %ymm14, %ymm12, %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm13
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm12
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm10, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm14
+; AVX512-FCP-NEXT: vpor %ymm12, %ymm14, %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm14
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm14, %ymm8
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm13
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm11
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm14
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm5
-; AVX512-FCP-NEXT: vmovdqa %ymm3, %ymm12
-; AVX512-FCP-NEXT: vporq %ymm5, %ymm0, %ymm21
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm2
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm11[0,1],ymm0[2],ymm11[3,4],ymm0[5],ymm11[6,7,8,9],ymm0[10],ymm11[11,12],ymm0[13],ymm11[14,15]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm11[0,1],ymm3[2],ymm11[3,4],ymm3[5],ymm11[6,7,8,9],ymm3[10],ymm11[11,12],ymm3[13],ymm11[14,15]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm16, %zmm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm17, %zmm3
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17,u,u,u,u],zero,zero
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm25, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm5
-; AVX512-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $248, %ymm12, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,4,0,0,0,5,0,0]
-; AVX512-FCP-NEXT: vpermd %ymm11, %ymm3, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm0, %ymm17, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm22
-; AVX512-FCP-NEXT: vprold $16, %ymm5, %ymm0
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm3, %ymm5, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm9
+; AVX512-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm8
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm9
+; AVX512-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm8, %ymm10
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm9
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm9, %ymm11
+; AVX512-FCP-NEXT: vporq %ymm11, %ymm10, %ymm22
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm12
+; AVX512-FCP-NEXT: vmovdqa64 %ymm13, %ymm30
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm11
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm12
+; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm12
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm7
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
+; AVX512-FCP-NEXT: vpternlogq $248, %ymm13, %ymm12, %ymm11
+; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm10
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm10, %ymm6
+; AVX512-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $248, %ymm16, %ymm11, %ymm6
+; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm12, %ymm11
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
+; AVX512-FCP-NEXT: vpermd %ymm7, %ymm12, %ymm12
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512-FCP-NEXT: vpternlogq $184, %ymm11, %ymm16, %ymm12
+; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm19
+; AVX512-FCP-NEXT: vprold $16, %ymm10, %ymm11
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $184, %ymm12, %ymm10, %ymm11
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm11[4,5,6,7]
+; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,1,4,5,4,5,5,7]
-; AVX512-FCP-NEXT: vmovdqa 96(%rax), %ymm0
-; AVX512-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vpandnq %ymm1, %ymm25, %ymm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
+; AVX512-FCP-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512-FCP-NEXT: vpermd %ymm6, %ymm11, %ymm11
+; AVX512-FCP-NEXT: vpandn %ymm11, %ymm13, %ymm11
+; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm18
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm6, %ymm12
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm11
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm12 = [151522058,0,421010202,421010202]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm4, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm23
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,2,2,3,10,9,11,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm4
+; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm12 = [218894094,0,488382238,488382238]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm11
+; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm16
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2,3],ymm11[4],ymm12[5,6,7,8],ymm11[9],ymm12[10,11],ymm11[12],ymm12[13,14,15]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm20, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm27, %zmm1
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm26
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5],ymm1[6],ymm3[7,8,9,10],ymm1[11],ymm3[12,13],ymm1[14],ymm3[15]
-; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm5 = [151522058,0,421010202,421010202]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2,3],ymm3[4],ymm6[5,6,7,8],ymm3[9],ymm6[10,11],ymm3[12],ymm6[13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,10,9,11,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
-; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm7 = [218894094,0,488382238,488382238]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm27
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm7[0],ymm1[1],ymm7[2,3],ymm1[4],ymm7[5,6,7,8],ymm1[9],ymm7[10,11],ymm1[12],ymm7[13,14,15]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm16
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm9[2],ymm7[3,4],ymm9[5],ymm7[6,7,8,9],ymm9[10],ymm7[11,12],ymm9[13],ymm7[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm7
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm1, %zmm7
-; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm24
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [5,0,0,0,6,0,0,6]
-; AVX512-FCP-NEXT: vpermd %ymm11, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpermd %ymm7, %ymm1, %ymm1
; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm12
+; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm1
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,30,0,0,0,31,0,0,31]
-; AVX512-FCP-NEXT: vpermi2d %zmm12, %zmm0, %zmm3
+; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm6, %zmm3
; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT: vpbroadcastd 72(%rax), %ymm0
-; AVX512-FCP-NEXT: vpandnq %ymm0, %ymm23, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512-FCP-NEXT: vpbroadcastd 72(%rax), %ymm3
+; AVX512-FCP-NEXT: vpandn %ymm3, %ymm6, %ymm3
; AVX512-FCP-NEXT: vmovdqa 64(%rax), %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm5
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm30
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm11
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm2
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,1,1,12,13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512-FCP-NEXT: vpternlogq $248, %zmm23, %zmm3, %zmm11
-; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
-; AVX512-FCP-NEXT: vpandnq %ymm0, %ymm23, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512-FCP-NEXT: vpternlogq $248, %zmm23, %zmm3, %zmm5
-; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm7 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm14, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3],ymm0[4,5],ymm3[6],ymm0[7,8,9,10],ymm3[11],ymm0[12,13],ymm3[14],ymm0[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm21, %zmm3
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm7, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm25
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm8, %ymm0
+; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm3
+; AVX512-FCP-NEXT: vpandn %ymm3, %ymm6, %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm4
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5],ymm0[6],ymm3[7,8,9,10],ymm0[11],ymm3[12,13],ymm0[14],ymm3[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm8, %ymm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm3
-; AVX512-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm22, %zmm4
+; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm3
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5],ymm3[6],ymm4[7,8,9,10],ymm3[11],ymm4[12,13],ymm3[14],ymm4[15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm4
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3],ymm4[4],ymm6[5,6,7,8],ymm4[9],ymm6[10,11],ymm4[12],ymm6[13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm4
+; AVX512-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm14, %ymm3
; AVX512-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm4
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm5
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm14, %ymm5
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm15[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 %zmm24, %zmm16
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm24, %zmm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm14, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vprold $16, %ymm14, %ymm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm13[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm19
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm0
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm27, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vprold $16, %ymm9, %ymm4
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm11, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm16
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm19, %zmm2
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm22, %zmm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm3
-; AVX512-FCP-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm5[2],ymm0[3,4],ymm5[5],ymm0[6,7,8,9],ymm5[10],ymm0[11,12],ymm5[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm1[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm4
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm5
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512-FCP-NEXT: vprold $16, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm2[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm14
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm6[3],ymm3[4,5],ymm6[6],ymm3[7,8,9,10],ymm6[11],ymm3[12,13],ymm6[14],ymm3[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm2, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm12, %zmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm25, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
+; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm16, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm4, %zmm19, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm10
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm10, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm19
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm17[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm9
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm10, %ymm4
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm8, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm21[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm4[2],ymm6[3,4],ymm4[5],ymm6[6,7,8,9],ymm4[10],ymm6[11,12],ymm4[13],ymm6[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm11, %zmm6
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm5, %zmm27, %zmm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm12
+; AVX512-FCP-NEXT: vprold $16, %ymm26, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm22
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm28[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm12, %ymm5
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm30
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm28[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
+; AVX512-FCP-NEXT: vpermd %zmm1, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm5, %zmm18, %zmm0
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm8, %ymm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm21[3,3,3,3,7,7,7,7]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm2, %xmm7
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [0,2,2,3,8,9,9,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm28[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm3
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm26
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm24
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm0
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm8, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm2
-; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm1
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm4 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm5
-; AVX512-FCP-NEXT: vmovdqa %xmm4, %xmm8
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512-FCP-NEXT: vmovdqa %xmm7, %xmm14
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm1, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm1, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm11
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm2
+; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm5 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm3
+; AVX512-FCP-NEXT: vmovdqa %xmm5, %xmm14
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm0, %zmm2
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm13
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm6
; AVX512-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vprold $16, %xmm6, %xmm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3,4],xmm1[5],xmm5[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm9, %xmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,1,1,8,8,10,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm5, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
+; AVX512-FCP-NEXT: vprold $16, %xmm6, %xmm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm28, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm5
+; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm6
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm2, %xmm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm6
; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,0,1,8,9,9,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm2
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm4, %xmm6
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm9[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm8[0],xmm6[1],xmm8[2,3],xmm6[4],xmm8[5,6],xmm6[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm8, %xmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm14, %xmm21
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm11, %zmm8
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vprold $16, %xmm11, %xmm6
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm4[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm28
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm27, %zmm9
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm8, %zmm5, %zmm9
-; AVX512-FCP-NEXT: vpbroadcastd 64(%rax), %ymm5
-; AVX512-FCP-NEXT: vpbroadcastd 68(%rax), %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm23
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm23
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,9,9,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm26, %zmm2
+; AVX512-FCP-NEXT: vpbroadcastd 64(%rax), %ymm6
+; AVX512-FCP-NEXT: vpbroadcastd 68(%rax), %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm31
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm21, %zmm31
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm23
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm5
-; AVX512-FCP-NEXT: vmovdqa64 %xmm10, %xmm22
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vmovdqa %xmm7, %xmm3
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm5, %zmm2, %zmm31
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm5
+; AVX512-FCP-NEXT: vmovdqa64 %xmm14, %xmm29
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm13, %zmm6
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512-FCP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vprold $16, %xmm13, %xmm5
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm0[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm28, %zmm8
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm3, %zmm8
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm3
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm5
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vmovdqa64 %xmm7, %xmm17
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm26, %zmm3
; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm5
; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm20
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm1, %zmm4, %zmm20
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm9, %zmm2, %zmm20
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm2
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm26
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm21, %zmm26
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm8, %zmm2, %zmm26
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm14, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm6[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vmovdqa %ymm6, %ymm9
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm3
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm5[2],ymm1[3,4],ymm5[5],ymm1[6,7,8,9],ymm5[10],ymm1[11,12],ymm5[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm9[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vmovdqa %ymm7, %ymm15
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm8
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm8
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512-FCP-NEXT: vprold $16, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm18
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm5
+; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %xmm11
+; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,8,9,9,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm8[0,1,2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8,9,10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm27, %zmm5
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512-FCP-NEXT: vprold $16, %ymm4, %ymm2
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm19[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm19[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
+; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm3
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm4
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm10
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm21, %zmm10
+; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm16, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
+; AVX512-FCP-NEXT: vpermd %zmm1, %zmm16, %zmm22
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm22
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm22
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,8,8,9]
+; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm6
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm5
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm2
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm14, %ymm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm9[0,1,2],ymm1[3],ymm9[4,5],ymm1[6],ymm9[7,8,9,10],ymm1[11],ymm9[12,13],ymm1[14],ymm9[15]
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm15
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm14
+; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm20, %zmm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [6,7,3,3,7,7,6,7]
+; AVX512-FCP-NEXT: vpermd %ymm25, %ymm17, %ymm16
+; AVX512-FCP-NEXT: vpbroadcastd 96(%rax), %ymm20
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm20, %zmm16, %zmm25
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm18, %zmm25
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm10, %zmm27, %zmm25
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512-FCP-NEXT: vprold $16, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm21, %zmm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,2,1,8,8,9,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm3
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm11, %xmm2
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0],xmm2[1],xmm7[2,3],xmm2[4],xmm7[5,6],xmm2[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm28, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm7
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm5[3],ymm2[4,5],ymm5[6],ymm2[7,8,9,10],ymm5[11],ymm2[12,13],ymm5[14],ymm2[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,2,3,8,8,8,9]
-; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm11
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm7
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm1
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm13
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm19, %zmm1
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512-FCP-NEXT: vpermd %zmm1, %zmm5, %zmm19
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm25, %zmm19
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm2
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm9[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm9[0],ymm2[1],ymm9[2,3],ymm2[4],ymm9[5,6,7,8],ymm2[9],ymm9[10,11],ymm2[12],ymm9[13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm31 = [6,7,3,3,7,7,6,7]
-; AVX512-FCP-NEXT: vpermd %ymm30, %ymm31, %ymm1
-; AVX512-FCP-NEXT: vpbroadcastd 96(%rax), %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm21
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm13, %zmm30, %zmm21
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm21
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm26, %xmm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm24[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0,1],xmm1[2],xmm2[3,4],xmm1[5],xmm2[6,7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm5
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm28, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm13, %xmm13
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [0,0,2,1,8,8,9,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm3
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq $184, %ymm0, %ymm18, %ymm16
+; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm0, %ymm13
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm3
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm11
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm2[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm30, %zmm8
+; AVX512-FCP-NEXT: vpbroadcastd 100(%rax), %ymm5
+; AVX512-FCP-NEXT: vpbroadcastd 104(%rax), %ymm19
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm5, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm6, %zmm24, %zmm5
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} ymm6 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm14
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %xmm12, %xmm30
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
+; AVX512-FCP-NEXT: vpermd %ymm23, %ymm17, %ymm17
+; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm19
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm17, %zmm20
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm8, %zmm18, %zmm20
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm1, %zmm27, %zmm20
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm7, %zmm1, %zmm5
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
+; AVX512-FCP-NEXT: vprold $16, %xmm9, %xmm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm15[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm8
; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm0
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm24 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm13
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0],xmm0[1],xmm9[2,3],xmm0[4],xmm9[5,6],xmm0[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm25, %zmm0
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm3, %ymm30, %ymm7
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm11
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm22, %xmm13
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm10, %xmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,1,1,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm28, %zmm3
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} ymm2 = ymm6[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm2[3,3,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm29 = ymm2[2,2,2,2]
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm1
-; AVX512-FCP-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512-FCP-NEXT: vpbroadcastd 104(%rax), %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm22, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm6[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,8,8,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm16, %zmm10
-; AVX512-FCP-NEXT: vpermd %ymm14, %ymm31, %ymm9
-; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm18
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm9, %zmm9
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm10, %zmm30, %zmm9
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm9
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm2
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm12, %xmm0
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0],xmm0[1],xmm8[2,3],xmm0[4],xmm8[5,6],xmm0[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm25, %zmm0
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm4
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm28, %zmm4
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm7
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm12[0],xmm7[1],xmm12[2,3],xmm7[4],xmm12[5,6],xmm7[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,0,1,1,8,8,10,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm0, %zmm7
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm8, %zmm28, %zmm7
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm3[4],xmm11[5],xmm3[5],xmm11[6],xmm3[6],xmm11[7],xmm3[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm0, %zmm2
; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm3
-; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm4, %zmm22, %zmm3
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm7, %ymm0, %ymm29
+; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm24, %zmm3
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm7, %zmm1, %zmm3
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm11
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm11, %ymm0, %ymm26
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm1
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm29[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm12
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd (%rsp), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512-FCP-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $184, %ymm13, %ymm1, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512-FCP-NEXT: vpternlogq $184, %ymm16, %ymm2, %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm2 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 # 64-byte Folded Reload
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: vpunpckhwd (%rsp), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512-FCP-NEXT: # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm5
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm7 = mem[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm8 = mem[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4],ymm8[5],ymm7[6,7,8,9],ymm8[10],ymm7[11,12],ymm8[13],ymm7[14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,4,5,4,5,5,7]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512-FCP-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17],zero,zero
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpandn %ymm8, %ymm11, %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512-FCP-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm9 = mem[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm10 = mem[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm13
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm10, %zmm25, %zmm1
+; AVX512-FCP-NEXT: vpermd %ymm11, %ymm10, %ymm10
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17],zero,zero
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm10
+; AVX512-FCP-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm11 = mem[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,3]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm14
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
; 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), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm4, %zmm25, %zmm5
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
-; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm6
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm11, %zmm28, %zmm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm28, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm7
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm10 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm10
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm20, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm23, 448(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 704(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 640(%rax)
-; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vmovaps %zmm0, 576(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 512(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 832(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 768(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm26, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 448(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 704(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 640(%rax)
+; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovaps %zmm1, 576(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 512(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 832(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 768(%rax)
; AVX512-FCP-NEXT: addq $1496, %rsp # imm = 0x5D8
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i16_stride7_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $2168, %rsp # imm = 0x878
-; AVX512DQ-NEXT: vmovdqa 96(%rcx), %ymm2
-; AVX512DQ-NEXT: vmovdqa 96(%rdx), %ymm9
+; AVX512DQ-NEXT: subq $2808, %rsp # imm = 0xAF8
+; AVX512DQ-NEXT: vmovdqa 96(%rcx), %ymm6
+; AVX512DQ-NEXT: vmovdqa 96(%rdx), %ymm13
; AVX512DQ-NEXT: vmovdqa 96(%rdi), %ymm7
; AVX512DQ-NEXT: vmovdqa 96(%rsi), %ymm8
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm9, %ymm2
-; AVX512DQ-NEXT: vporq %ymm1, %ymm2, %ymm16
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm8, %ymm1
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm7, %ymm2
-; AVX512DQ-NEXT: vporq %ymm1, %ymm2, %ymm17
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm6, %ymm2
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm3
+; AVX512DQ-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm8, %ymm2
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm7, %ymm3
+; AVX512DQ-NEXT: vporq %ymm2, %ymm3, %ymm18
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm12 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
; AVX512DQ-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa %ymm3, %ymm10
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm2, %ymm2
; AVX512DQ-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm15 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm3, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm20
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm6
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm6, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 64(%rsi), %ymm5
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm5, %ymm1
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm3, %ymm3
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm27
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%rsi), %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm23
; AVX512DQ-NEXT: vmovdqa 64(%rdi), %ymm4
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm4, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm1, %ymm1
-; AVX512DQ-NEXT: vmovdqa %ymm10, %ymm3
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm4, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm22
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm2, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm1, %ymm1
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm2, %ymm2
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm3
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm3, %ymm3
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm2
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm3, %ymm3
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm2, %ymm2
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm10, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm10, %ymm22
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm0
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm1, %ymm11
-; AVX512DQ-NEXT: vpor %ymm0, %ymm11, %ymm0
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm2, %ymm2
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX512DQ-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm0, %ymm12
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm11, %ymm13
-; AVX512DQ-NEXT: vpor %ymm12, %ymm13, %ymm10
-; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm15, %ymm13, %ymm12
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm15, %ymm14
-; AVX512DQ-NEXT: vpor %ymm14, %ymm12, %ymm10
-; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vprold $16, %ymm15, %ymm12
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm13[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0,1],ymm12[2],ymm14[3,4],ymm12[5],ymm14[6,7,8,9],ymm12[10],ymm14[11,12],ymm12[13],ymm14[14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm14 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm5, %ymm0
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm3, %ymm1
+; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm1, %ymm9
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm0, %ymm10
+; AVX512DQ-NEXT: vpor %ymm10, %ymm9, %ymm9
+; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm10, %ymm6, %ymm9
+; AVX512DQ-NEXT: vmovdqa64 %ymm10, %ymm31
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm9 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm10 = ymm14[0,1],ymm10[2],ymm14[3,4],ymm10[5],ymm14[6,7,8,9],ymm10[10],ymm14[11,12],ymm10[13],ymm14[14,15]
-; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-NEXT: vpermi2q %zmm10, %zmm12, %zmm14
-; AVX512DQ-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm14
-; AVX512DQ-NEXT: vpshufb %ymm12, %ymm14, %ymm10
-; AVX512DQ-NEXT: vmovdqa64 %ymm12, %ymm19
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1],ymm12[2],ymm10[3,4],ymm12[5],ymm10[6,7,8,9],ymm12[10],ymm10[11,12],ymm12[13],ymm10[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2,3],ymm10[4],ymm12[5,6,7,8],ymm10[9],ymm12[10,11],ymm10[12],ymm12[13,14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm10 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm9 = ymm8[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm9 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0,1,2],ymm9[3],ymm10[4,5],ymm9[6],ymm10[7,8,9,10],ymm9[11],ymm10[12,13],ymm9[14],ymm10[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,3,3]
+; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 96(%r9), %ymm9
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm9[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
; AVX512DQ-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm14[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,3,2]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm9[2,1,2,3,6,5,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm10[2,2,2,2]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm6 = ymm10[0,1],ymm6[2],ymm10[3,4],ymm6[5],ymm10[6,7,8,9],ymm6[10],ymm10[11,12],ymm6[13],ymm10[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[1,1,1,1,5,5,5,5]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,2,1,4,4,6,5]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4],ymm7[5],ymm8[6,7,8,9],ymm7[10],ymm8[11,12],ymm7[13],ymm8[14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm16, %zmm8
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm17, %zmm7
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm7
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vmovdqa 96(%r8), %ymm10
+; AVX512DQ-NEXT: vprold $16, %ymm9, %ymm8
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm9, %ymm9
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm17, %zmm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm7
+; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
+; AVX512DQ-NEXT: vmovdqa 96(%r8), %ymm6
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[1,2,2,3,5,6,6,7]
; AVX512DQ-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm8 = ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17,u,u,u,u],zero,zero
-; AVX512DQ-NEXT: vpternlogq $248, %ymm26, %ymm7, %ymm8
-; AVX512DQ-NEXT: vmovdqa 96(%r9), %ymm12
-; AVX512DQ-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm12, %ymm9
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $248, %ymm14, %ymm8, %ymm9
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm7, %ymm7
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm8 = ymm10[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[16,17,u,u,u,u],zero,zero
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $248, %ymm11, %ymm7, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm12
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $248, %ymm11, %ymm6, %ymm9
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm7, %ymm6
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm10[2,1,3,3]
; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm7, %ymm10, %ymm8
-; AVX512DQ-NEXT: vprold $16, %ymm12, %ymm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,2]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm8, %ymm14, %ymm7
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm9[0,1,2,3],zmm7[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpternlogq $184, %ymm6, %ymm10, %ymm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm8[2,2,2,2]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %ymm7, %ymm8, %ymm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm9[0,1,2,3],zmm6[4,5,6,7]
+; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,5,0,0,0,6,0,0,6,0,0,0,7,0,0,7]
-; AVX512DQ-NEXT: vmovdqa 96(%rax), %ymm7
-; AVX512DQ-NEXT: vpermd %zmm7, %zmm18, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm10 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm7, %ymm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
-; AVX512DQ-NEXT: vpandnq %ymm9, %ymm26, %ymm9
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512DQ-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512DQ-NEXT: vpermd %zmm6, %zmm18, %zmm7
; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-NEXT: vpbroadcastd 72(%rax), %ymm7
-; AVX512DQ-NEXT: vpandnq %ymm7, %ymm16, %ymm9
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm6, %ymm6
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,2,3]
+; AVX512DQ-NEXT: vpandn %ymm7, %ymm12, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm19
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512DQ-NEXT: vpbroadcastd 72(%rax), %ymm6
+; AVX512DQ-NEXT: vpandnq %ymm6, %ymm28, %ymm6
; AVX512DQ-NEXT: vmovdqa 64(%rax), %ymm7
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm7, %ymm12
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpbroadcastd 8(%rax), %ymm9
-; AVX512DQ-NEXT: vpandnq %ymm9, %ymm16, %ymm9
-; AVX512DQ-NEXT: vmovdqa (%rax), %ymm8
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm8, %ymm12
-; AVX512DQ-NEXT: vmovdqa64 %ymm8, %ymm24
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
-; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0,1],ymm9[2],ymm12[3,4],ymm9[5],ymm12[6,7,8,9],ymm9[10],ymm12[11,12],ymm9[13],ymm12[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm12[2],ymm9[3,4],ymm12[5],ymm9[6,7,8,9],ymm12[10],ymm9[11,12],ymm12[13],ymm9[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm9, %ymm10
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm7, %ymm7
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%r9), %xmm7
+; AVX512DQ-NEXT: vmovdqa 64(%r8), %xmm8
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm17
+; AVX512DQ-NEXT: vmovdqa64 %xmm7, %xmm20
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-NEXT: vpshufb %xmm7, %xmm6, %xmm6
+; AVX512DQ-NEXT: vmovdqa %xmm7, %xmm15
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%rcx), %xmm9
+; AVX512DQ-NEXT: vmovdqa 64(%rdx), %xmm7
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
+; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm14
+; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm10
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpbroadcastd 8(%rax), %ymm8
+; AVX512DQ-NEXT: vpandnq %ymm8, %ymm28, %ymm8
+; AVX512DQ-NEXT: vmovdqa (%rax), %ymm12
+; AVX512DQ-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm12, %ymm13
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
+; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm6
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm12
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm12, %xmm29
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm24
+; AVX512DQ-NEXT: vpshufb %xmm15, %xmm8, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm15, %xmm25
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,1]
+; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm8
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm13
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,0,2,1]
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm12, %xmm21
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm16
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm15[0,0,1,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm15 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm12 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rax), %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm15, %ymm11
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm15[0,1,1,3,4,5,5,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,2,2,3]
-; AVX512DQ-NEXT: vpandnq %ymm12, %ymm26, %ymm12
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10
-; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm13[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm12[0,1,2],ymm10[3],ymm12[4,5],ymm10[6],ymm12[7,8,9,10],ymm10[11],ymm12[12,13],ymm10[14],ymm12[15]
-; AVX512DQ-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm2, %ymm10
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0,1],ymm12[2],ymm10[3,4],ymm12[5],ymm10[6,7,8,9],ymm12[10],ymm10[11,12],ymm12[13],ymm10[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpandnq %ymm12, %ymm19, %ymm12
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
+; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm11 = ymm0[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1,2],ymm12[3],ymm11[4,5],ymm12[6],ymm11[7,8,9,10],ymm12[11],ymm11[12,13],ymm12[14],ymm11[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,3,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm31, %ymm6
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm4, %ymm11
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm12[2],ymm11[3,4],ymm12[5],ymm11[6,7,8,9],ymm12[10],ymm11[11,12],ymm12[13],ymm11[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6,7,8],ymm4[9],ymm2[10,11],ymm4[12],ymm2[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm2 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6,7,8],ymm2[9],ymm1[10,11],ymm2[12],ymm1[13,14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm31
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm3 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,3]
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vprold $16, %ymm0, %ymm2
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpermd %zmm9, %zmm18, %zmm0
+; AVX512DQ-NEXT: vpermd %zmm15, %zmm18, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm12
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[1,1,1,1,5,5,5,5]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm22[1,1,1,1,5,5,5,5]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm12[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm22[2,2,2,2,6,6,6,6]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm2
+; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm2
; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm27[0,1,1,3,4,5,5,7]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm30
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm2, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm3
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm31, %ymm11
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm27[2,2,2,2,6,6,6,6]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm8
-; AVX512DQ-NEXT: vprold $16, %ymm21, %ymm0
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm20[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm10, %zmm1
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm11 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512DQ-NEXT: vpermd 64(%rax), %zmm11, %zmm0
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm26, %zmm0
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vprold $16, %ymm5, %ymm0
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,2,2,3,5,6,6,7]
+; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm6
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm27[3,3,3,3,7,7,7,7]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5,6,7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13,14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm8[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm12 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 96(%rcx), %xmm0
+; AVX512DQ-NEXT: vmovdqa 96(%rdx), %xmm1
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm3 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7,8,9,10],ymm3[11],ymm2[12,13],ymm3[14],ymm2[15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,3,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 96(%rsi), %xmm2
+; AVX512DQ-NEXT: vmovdqa 96(%rdi), %xmm3
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm12
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm4, %xmm4
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vprold $16, %xmm2, %xmm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
+; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm4
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[1,1,2,2]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vmovdqa 96(%r9), %xmm0
; AVX512DQ-NEXT: vmovdqa 96(%r8), %xmm1
; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpermt2d %zmm4, %zmm12, %zmm3
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[2,3,3,3,6,7,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,3,2]
-; AVX512DQ-NEXT: vpbroadcastd 96(%rax), %ymm5
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm3, %zmm5, %zmm4
-; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm15
-; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 96(%rsi), %xmm3
-; AVX512DQ-NEXT: vmovdqa 96(%rdi), %xmm4
-; AVX512DQ-NEXT: vprold $16, %xmm3, %xmm5
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
-; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm23
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512DQ-NEXT: vmovdqa 96(%rdx), %xmm4
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm6 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm3, %xmm5
-; AVX512DQ-NEXT: vmovdqa %xmm6, %xmm7
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,7,6]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm29 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
-; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm29, %zmm1
-; AVX512DQ-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512DQ-NEXT: vpbroadcastd 104(%rax), %ymm3
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm14, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqa %xmm7, %xmm8
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
+; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm18 = [16,18,19,19,19,19,0,0,0,1,0,1,2,3,2,3]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm18, %zmm0
+; AVX512DQ-NEXT: vpbroadcastd 100(%rax), %ymm1
+; AVX512DQ-NEXT: vpbroadcastd 104(%rax), %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm31
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm31
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
+; AVX512DQ-NEXT: vpshufb %xmm4, %xmm9, %xmm1
+; AVX512DQ-NEXT: vmovdqa %xmm4, %xmm6
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[1,1,2,2]
; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [16,16,17,17,17,17,0,0,0,1,0,1,2,3,2,3]
-; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm3, %zmm1
-; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm2
-; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm4
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm5, %xmm5
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,1]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm25
-; AVX512DQ-NEXT: vprold $16, %xmm4, %xmm4
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm4[2],xmm2[3,4],xmm4[5],xmm2[6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 64(%r9), %xmm1
-; AVX512DQ-NEXT: vmovdqa 64(%r8), %xmm2
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm17
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,7,6]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
-; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm6, %zmm1
-; AVX512DQ-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512DQ-NEXT: vpbroadcastd 68(%rax), %ymm5
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm9, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm5
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512DQ-NEXT: vpshufb %xmm8, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqa %xmm8, %xmm14
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,2]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm5[0],xmm1[1],xmm5[2,3],xmm1[4],xmm5[5,6],xmm1[7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm7[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm3, %zmm1
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm19
-; AVX512DQ-NEXT: vprold $16, %xmm5, %xmm5
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm5[2],xmm3[3,4],xmm5[5],xmm3[6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm2 = [16,16,17,17,17,17,0,0,0,1,0,1,2,3,2,3]
+; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm2, %zmm1
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
+; AVX512DQ-NEXT: vprold $16, %xmm10, %xmm3
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm4 = xmm14[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm2
-; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm3
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm14
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm0, %zmm14
+; AVX512DQ-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm3
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,7,6]
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm6, %zmm1
-; AVX512DQ-NEXT: vpbroadcastd (%rax), %ymm3
-; AVX512DQ-NEXT: vpbroadcastd 4(%rax), %ymm4
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm2
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm9, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm28, %ymm7
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm22[1,1,1,1,5,5,5,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm22
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm7[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm9
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2,3],ymm1[4],ymm4[5,6,7,8],ymm1[9],ymm4[10,11],ymm1[12],ymm4[13,14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm21
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,1,0,1,0,1,1,3,16,18,19,19,19,19,0,0]
+; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm4, %zmm1
+; AVX512DQ-NEXT: vpbroadcastd 64(%rax), %ymm3
+; AVX512DQ-NEXT: vpbroadcastd 68(%rax), %ymm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm25
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm19, %zmm25
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3]
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm8, %xmm3
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm13[1,1,2,2]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm2, %zmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm16, %xmm5
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm2
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm1, %xmm1
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,1]
+; AVX512DQ-NEXT: vprold $16, %xmm21, %xmm2
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm16[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm5[0,1],xmm2[2],xmm5[3,4],xmm2[5],xmm5[6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm5
+; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm0, %zmm5
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,7,6]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm4, %zmm0
+; AVX512DQ-NEXT: vpbroadcastd (%rax), %ymm1
+; AVX512DQ-NEXT: vpbroadcastd 4(%rax), %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm20
+; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm19, %zmm20
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm10
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-NEXT: vprold $16, %ymm2, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm4 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm3[0,0,2,1,4,4,6,5]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm6[3],ymm4[4,5],ymm6[6],ymm4[7,8,9,10],ymm6[11],ymm4[12,13],ymm6[14],ymm4[15]
-; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm10, %zmm4
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm6[0,1],ymm1[2],ymm6[3,4],ymm1[5],ymm6[6,7,8,9],ymm1[10],ymm6[11,12],ymm1[13],ymm6[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm8, %ymm1
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm6[2],ymm1[3,4],ymm6[5],ymm1[6,7,8,9],ymm6[10],ymm1[11,12],ymm6[13],ymm1[14,15]
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm18
-; AVX512DQ-NEXT: vpermd (%rax), %zmm11, %zmm1
-; AVX512DQ-NEXT: vpternlogd $184, %zmm4, %zmm26, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2,3],ymm4[4],ymm1[5,6,7,8],ymm4[9],ymm1[10,11],ymm4[12],ymm1[13,14,15]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm7[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm4[3],ymm1[4,5],ymm4[6],ymm1[7,8,9,10],ymm4[11],ymm1[12,13],ymm4[14],ymm1[15]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm4 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm11 = xmm6[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm12, %zmm1
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm11 = ymm24[2,3,3,3,6,7,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512DQ-NEXT: vpbroadcastd 32(%rax), %ymm12
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm26
-; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm15, %zmm26
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512DQ-NEXT: vprold $16, %xmm12, %xmm15
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm13 = xmm1[1,1,2,3]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm2 = xmm13[0,1],xmm15[2],xmm13[3,4],xmm15[5],xmm13[6,7]
-; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm12[0],xmm1[1],xmm12[1],xmm1[2],xmm12[2],xmm1[3],xmm12[3]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm12[4],xmm1[4],xmm12[5],xmm1[5],xmm12[6],xmm1[6],xmm12[7],xmm1[7]
-; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm17, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm5, %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm23, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm13, %xmm1
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm4, %xmm4
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm24 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm23 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm7 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm17 = mem[2,1,3,3]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm6[0,1,2,3,4,5,7,6]
-; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm29, %zmm4
-; AVX512DQ-NEXT: vpbroadcastd 36(%rax), %ymm0
-; AVX512DQ-NEXT: vpbroadcastd 40(%rax), %ymm6
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm28
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm28
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm6, %xmm4
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
-; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm12[0],xmm4[1],xmm12[2,3],xmm4[4],xmm12[5,6],xmm4[7]
-; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
-; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm29 = ymm0[3,3,3,3]
-; AVX512DQ-NEXT: vpshufd $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm0 = mem[1,2,2,3,5,6,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm1[2,1,2,3,6,5,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm10 = xmm1[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm1[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm15 = xmm1[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
-; AVX512DQ-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm12 = xmm1[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,3]
-; AVX512DQ-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm4 = mem[2,1,3,2]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm5 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm6 = mem[2,2,3,3]
-; AVX512DQ-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm1 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm1 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,2,2,3]
-; AVX512DQ-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm27 = mem[2,1,3,3]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm20 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm19 = ymm30[2,1,3,2]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm1 = mem[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm25 = ymm22[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm21[0,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,1,3,2]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm18[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm23
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm7, %zmm24
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %zmm23, %zmm30, %zmm24
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[0,1,1,3,4,5,5,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm24, %ymm7, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm0, %ymm18, %ymm14
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm10, %zmm23, %zmm8
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm17 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm17, %zmm8
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm10 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm23, %zmm10
-; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm17, %zmm10
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm2, %ymm0
+; AVX512DQ-NEXT: vmovdqa %ymm2, %ymm9
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vmovdqa %ymm3, %ymm11
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[0,0,2,1,4,4,6,5]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm0[0,1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8,9,10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
+; AVX512DQ-NEXT: vprold $16, %ymm7, %ymm0
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[1,2,2,3,5,6,6,7]
+; AVX512DQ-NEXT: vmovdqa %ymm4, %ymm8
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7,8,9],ymm0[10],ymm2[11,12],ymm0[13],ymm2[14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm2 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6,7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm10[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm6[3],ymm0[4,5],ymm6[6],ymm0[7,8,9,10],ymm6[11],ymm0[12,13],ymm6[14],ymm0[15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
+; AVX512DQ-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm22 = ymm1[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm26 = ymm3[2,2,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm24 = ymm4[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-NEXT: vprold $16, %xmm10, %xmm1
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm9[1,1,2,3]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2],xmm3[3,4],xmm1[5],xmm3[6,7]
+; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm21 = ymm2[0,2,2,3]
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm2
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm2, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm1, %xmm1
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,4,5,7,6]
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm27
+; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm18, %zmm1
+; AVX512DQ-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512DQ-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm13
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm13
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm7
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm12 = xmm4[0,0,1,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm0[2,1,3,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm3[0,0,1,1]
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm2
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm21
-; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm24, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm2, %ymm3, %ymm29
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm22 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm3, %zmm4, %zmm1
+; AVX512DQ-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm3 = mem[2,1,3,2]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm29 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %ymm1, %ymm29, %ymm3
+; AVX512DQ-NEXT: vpternlogq $184, %ymm3, %ymm28, %ymm30
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm0[0,1,2,3],zmm6[0,1,2,3]
-; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm24 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm2
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm14[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm0, %zmm29
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm18 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm17 # 64-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm20, %zmm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm19, %zmm3
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm3
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm25, %zmm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm16, %zmm5
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm4, %zmm5
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, %zmm3, %zmm2, %zmm20
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, %zmm5, %zmm2, %zmm25
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm3 # 64-byte Folded Reload
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm27, %zmm31, %zmm3
-; AVX512DQ-NEXT: vpternlogq $184, %zmm2, %zmm30, %zmm3
-; AVX512DQ-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm4 = mem[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,1,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512DQ-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm4 = mem[2,1,3,3]
-; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm5 = mem[0,0,1,1]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm5 = mem[0,0,2,1]
-; AVX512DQ-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm6 = mem[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
-; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm19 = mem[0,0,1,1]
-; AVX512DQ-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm9 = mem[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512DQ-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512DQ-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
-; AVX512DQ-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm13 = mem[2,1,3,3]
-; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm14 = mem[0,0,1,1]
-; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm15 = mem[0,0,2,1]
-; AVX512DQ-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm0 = mem[2,1,2,3,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,3]
-; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm16 = mem[0,0,1,1]
-; AVX512DQ-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm1 = mem[0,2,3,3,4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm30, %zmm4
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm12[0,1,1,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm11, %zmm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm11
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm30, %zmm11
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, %zmm4, %zmm2, %zmm12
-; AVX512DQ-NEXT: vpternlogq $184, %zmm11, %zmm2, %zmm26
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm19, %zmm4
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm23, %zmm4
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm16, %zmm1
-; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm23, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm23 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm18 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm23, %zmm28, %zmm18
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm18
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm23 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm23
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm23
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm17 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm17
+; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512DQ-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm19, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-NEXT: vpternlogq $184, %ymm0, %ymm1, %ymm2
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm19 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm1
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm30[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Reload
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm30
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm11
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm14, %zmm1, %zmm25
+; AVX512DQ-NEXT: vpternlogq $184, %zmm5, %zmm1, %zmm20
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogq $184, %zmm4, %zmm0, %zmm2
-; AVX512DQ-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm28
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm24
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm7
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm5, %zmm2
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm22, %zmm3, %zmm22
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm5, %zmm22
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-NEXT: vpermd 64(%rax), %zmm14, %zmm5
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm3, %zmm5
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm24, %zmm26, %zmm1
+; AVX512DQ-NEXT: vpermd (%rax), %zmm14, %zmm14
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm3, %zmm14
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm5
+; AVX512DQ-NEXT: vpternlogq $184, %zmm22, %zmm1, %zmm14
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $184, %zmm1, %zmm4, %zmm2
+; AVX512DQ-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm1 = mem[0,1,1,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm22 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm22
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm12[0,1,1,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm21, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm15, %zmm8
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm8
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm4 = xmm7[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[1,1,2,2]
+; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm4 = xmm9[0],xmm4[1],xmm9[2,3],xmm4[4],xmm9[5,6],xmm4[7]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm7 = mem[2,2,2,3]
+; AVX512DQ-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm9 = mem[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[0,1,1,3]
+; AVX512DQ-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm10 = mem[2,3,3,3,6,7,7,7]
+; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm16 = mem[0,0,2,1]
+; AVX512DQ-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm12 = mem[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,3]
+; AVX512DQ-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm21 = mem[0,0,1,1]
+; AVX512DQ-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm15 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm24 = mem[2,2,2,3]
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm3
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,1,3]
+; AVX512DQ-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm26 = mem[2,3,3,3,6,7,7,7]
+; AVX512DQ-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm27 = mem[0,0,2,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,1,2,3,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,4]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm10[2,1,3,2]
+; AVX512DQ-NEXT: vpbroadcastd 96(%rax), %ymm10
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm24, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm26[2,1,3,2]
+; AVX512DQ-NEXT: vpbroadcastd 32(%rax), %ymm10
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm7
+; AVX512DQ-NEXT: vpternlogd $184, %zmm3, %zmm29, %zmm7
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm22, %zmm3, %zmm9
+; AVX512DQ-NEXT: vpternlogq $184, %zmm8, %zmm3, %zmm7
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm16, %zmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm8
+; AVX512DQ-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm8
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm27, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm3
+; AVX512DQ-NEXT: vpternlogq $226, %zmm1, %zmm28, %zmm3
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm8, %zmm1, %zmm31
+; AVX512DQ-NEXT: vpternlogq $184, %zmm3, %zmm1, %zmm13
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm0
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm11
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm24, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm28, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm26, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm25, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm17, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm18, 448(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 704(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm12, 640(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm20, 576(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm8, 512(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm22, 768(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm29, 832(%rax)
-; AVX512DQ-NEXT: addq $2168, %rsp # imm = 0x878
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm13, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm14, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm23, 64(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm20, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm25, 448(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm31, 704(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm9, 640(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm5, 576(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm18, 512(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm11, 384(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm19, 768(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm30, 832(%rax)
+; AVX512DQ-NEXT: addq $2808, %rsp # imm = 0xAF8
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i16_stride7_vf64:
; AVX512DQ-FCP: # %bb.0:
; AVX512DQ-FCP-NEXT: subq $1496, %rsp # imm = 0x5D8
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm10
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm16
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm2
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %ymm5
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm28
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm4
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm4, %ymm18
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm31
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm12
-; AVX512DQ-FCP-NEXT: vpor %ymm8, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm17
+; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm13, %ymm19
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm15
+; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm12
+; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm12, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm13
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm10
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm10, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm14
+; AVX512DQ-FCP-NEXT: vpor %ymm12, %ymm14, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm14
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm14, %ymm8
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm13
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm11
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm14
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, %ymm12
-; AVX512DQ-FCP-NEXT: vporq %ymm5, %ymm0, %ymm21
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm2
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm11[0,1],ymm0[2],ymm11[3,4],ymm0[5],ymm11[6,7,8,9],ymm0[10],ymm11[11,12],ymm0[13],ymm11[14,15]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm11[0,1],ymm3[2],ymm11[3,4],ymm3[5],ymm11[6,7,8,9],ymm3[10],ymm11[11,12],ymm3[13],ymm11[14,15]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm16, %zmm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm17, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17,u,u,u,u],zero,zero
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm25, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm12, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,4,0,0,0,5,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm3, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm0, %ymm17, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm22
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm3, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm9
+; AVX512DQ-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm9
+; AVX512DQ-FCP-NEXT: vpor %ymm8, %ymm9, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm8, %ymm10
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm9, %ymm11
+; AVX512DQ-FCP-NEXT: vporq %ymm11, %ymm10, %ymm22
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm13, %ymm30
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm11
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17,u,u,u,u],zero,zero
+; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm13, %ymm12, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm10, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $248, %ymm16, %ymm11, %ymm6
+; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm12, %ymm11
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm12, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm11, %ymm16, %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm19
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm10, %ymm11
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm12, %ymm10, %ymm11
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm11[4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rax), %ymm0
-; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vpandnq %ymm1, %ymm25, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512DQ-FCP-NEXT: vpermd %ymm6, %ymm11, %ymm11
+; AVX512DQ-FCP-NEXT: vpandn %ymm11, %ymm13, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm18
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm6, %ymm12
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm11
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm12 = [151522058,0,421010202,421010202]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3],ymm4[4],ymm5[5,6,7,8],ymm4[9],ymm5[10,11],ymm4[12],ymm5[13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,2,2,3,10,9,11,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm12 = [218894094,0,488382238,488382238]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2,3],ymm11[4],ymm12[5,6,7,8],ymm11[9],ymm12[10,11],ymm11[12],ymm12[13,14,15]
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm27, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm26
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5],ymm1[6],ymm3[7,8,9,10],ymm1[11],ymm3[12,13],ymm1[14],ymm3[15]
-; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm5 = [151522058,0,421010202,421010202]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2,3],ymm3[4],ymm6[5,6,7,8],ymm3[9],ymm6[10,11],ymm3[12],ymm6[13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,10,9,11,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm7 = [218894094,0,488382238,488382238]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm27
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm7[0],ymm1[1],ymm7[2,3],ymm1[4],ymm7[5,6,7,8],ymm1[9],ymm7[10,11],ymm1[12],ymm7[13,14,15]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm16
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm9[2],ymm7[3,4],ymm9[5],ymm7[6,7,8,9],ymm9[10],ymm7[11,12],ymm9[13],ymm7[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm1, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm24
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [5,0,0,0,6,0,0,6]
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm1
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,5,0,0,0,6,0,0,30,0,0,0,31,0,0,31]
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm12, %zmm0, %zmm3
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm6, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT: vpbroadcastd 72(%rax), %ymm0
-; AVX512DQ-FCP-NEXT: vpandnq %ymm0, %ymm23, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512DQ-FCP-NEXT: vpbroadcastd 72(%rax), %ymm3
+; AVX512DQ-FCP-NEXT: vpandn %ymm3, %ymm6, %ymm3
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rax), %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm30
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm11
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm2
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [0,0,1,1,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm23, %zmm3, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
-; AVX512DQ-FCP-NEXT: vpandnq %ymm0, %ymm23, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm23, %zmm3, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} ymm7 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm14, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm3[3],ymm0[4,5],ymm3[6],ymm0[7,8,9,10],ymm3[11],ymm0[12,13],ymm3[14],ymm0[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm21, %zmm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm25
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm3
+; AVX512DQ-FCP-NEXT: vpandn %ymm3, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5],ymm0[6],ymm3[7,8,9,10],ymm0[11],ymm3[12,13],ymm0[14],ymm3[15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} ymm4 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm22, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5],ymm3[6],ymm4[7,8,9,10],ymm3[11],ymm4[12,13],ymm3[14],ymm4[15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3],ymm4[4],ymm6[5,6,7,8],ymm4[9],ymm6[10,11],ymm4[12],ymm6[13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm14, %ymm3
; AVX512DQ-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm5
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6,7,8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm14, %ymm5
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm15[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, %zmm16
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm24, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm14, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm14, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm13[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm18 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm19
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7,8,9],ymm6[10],ymm5[11,12],ymm6[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm27, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3,4],ymm4[5],ymm5[6,7,8,9],ymm4[10],ymm5[11,12],ymm4[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm11, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm16
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,21,0,0,0,22,0,0,14,0,0,0,15,0,0,15]
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm19, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm22, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm5[2],ymm0[3,4],ymm5[5],ymm0[6,7,8,9],ymm5[10],ymm0[11,12],ymm5[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm1[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm4
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm18, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm2[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm14
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm6[3],ymm3[4,5],ymm6[6],ymm3[7,8,9,10],ymm6[11],ymm3[12,13],ymm6[14],ymm3[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm2, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm12, %zmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm25, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm16, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm4, %zmm19, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm10, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm19
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm17[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6,7,8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm10, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm20, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm8, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm21[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm4[2],ymm6[3,4],ymm4[5],ymm6[6,7,8,9],ymm4[10],ymm6[11,12],ymm4[13],ymm6[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm11, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm5, %zmm27, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm12
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm26, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm22
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm28[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4],ymm3[5],ymm5[6,7,8,9],ymm3[10],ymm5[11,12],ymm3[13],ymm5[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm12, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm30
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm28[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm5
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm25, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm5, %zmm18, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm21[3,3,3,3,7,7,7,7]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, %xmm7
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [0,2,2,3,8,9,9,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm28[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm26
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm24
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm8, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %xmm1
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm4 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, %xmm8
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm1, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm1, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %xmm2
+; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm5 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, %xmm14
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm6
; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm6, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3,4],xmm1[5],xmm5[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, %xmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [0,0,1,1,8,8,10,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm5, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3]
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm6, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0,1],xmm3[2],xmm5[3,4],xmm3[5],xmm5[6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm5, %xmm5
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm28, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm6
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm2, %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm2, %xmm6
; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm4, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm9[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm8[0],xmm6[1],xmm8[2,3],xmm6[4],xmm8[5,6],xmm6[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm8, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm14, %xmm21
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm11, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm11, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm4[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm28
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm27, %zmm9
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm8, %zmm5, %zmm9
-; AVX512DQ-FCP-NEXT: vpbroadcastd 64(%rax), %ymm5
-; AVX512DQ-FCP-NEXT: vpbroadcastd 68(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm23
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm23
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm26, %zmm2
+; AVX512DQ-FCP-NEXT: vpbroadcastd 64(%rax), %ymm6
+; AVX512DQ-FCP-NEXT: vpbroadcastd 68(%rax), %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm31
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm21, %zmm31
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm23
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm10, %xmm22
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, %xmm3
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm5, %zmm2, %zmm31
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm14, %xmm29
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm13, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm13, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm0[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm8[0,1],xmm5[2],xmm8[3,4],xmm5[5],xmm8[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm28, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm3, %zmm8
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm7, %xmm17
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm26, %zmm3
; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm5
; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm20
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm1, %zmm4, %zmm20
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm9, %zmm2, %zmm20
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm26
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm21, %zmm26
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm8, %zmm2, %zmm26
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm14, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm6[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, %ymm9
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm5[2],ymm3[3,4],ymm5[5],ymm3[6,7,8,9],ymm5[10],ymm3[11,12],ymm5[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm5[2],ymm1[3,4],ymm5[5],ymm1[6,7,8,9],ymm5[10],ymm1[11,12],ymm5[13],ymm1[14,15]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm9[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, %ymm15
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm18, %zmm8
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm18
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7,8,9],ymm5[10],ymm6[11,12],ymm5[13],ymm6[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm6 = [0,2,2,3,8,9,9,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm8[0,1,2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8,9,10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm27, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm4, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm19[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm19[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm3[0,1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7,8,9,10],ymm4[11],ymm3[12,13],ymm4[14],ymm3[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm4
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm20, %zmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm21, %zmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm16, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm16, %zmm22
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm22
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm22
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6,7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm30 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm5
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm14, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm9[0,1,2],ymm1[3],ymm9[4,5],ymm1[6],ymm9[7,8,9,10],ymm1[11],ymm9[12,13],ymm1[14],ymm9[15]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm14
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [6,7,3,3,7,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm25, %ymm17, %ymm16
+; AVX512DQ-FCP-NEXT: vpbroadcastd 96(%rax), %ymm20
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm20, %zmm16, %zmm25
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm18, %zmm25
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm10, %zmm27, %zmm25
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1],xmm3[2],xmm4[3,4],xmm3[5],xmm4[6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm21, %zmm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm2[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,2,1,8,8,9,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm3
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm11, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0],xmm2[1],xmm7[2,3],xmm2[4],xmm7[5,6],xmm2[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm28, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm28, %zmm7
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm18
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm5[3],ymm2[4,5],ymm5[6],ymm2[7,8,9,10],ymm5[11],ymm2[12,13],ymm5[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm7
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm1
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm13
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm19, %zmm1
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,0,4,0,0,0,5,0,0,13,0,0,0,14,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm5, %zmm19
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm25, %zmm19
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm2
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm9[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm9[0],ymm2[1],ymm9[2,3],ymm2[4],ymm9[5,6,7,8],ymm2[9],ymm9[10,11],ymm2[12],ymm9[13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm31 = [6,7,3,3,7,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm30, %ymm31, %ymm1
-; AVX512DQ-FCP-NEXT: vpbroadcastd 96(%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm21
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm13, %zmm30, %zmm21
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm21
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm26, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm24[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0,1],xmm1[2],xmm2[3,4],xmm1[5],xmm2[6,7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm13, %xmm13
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm17 = [0,0,2,1,8,8,9,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm0, %ymm18, %ymm16
+; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm0, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm11
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm2[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm30, %zmm8
+; AVX512DQ-FCP-NEXT: vpbroadcastd 100(%rax), %ymm5
+; AVX512DQ-FCP-NEXT: vpbroadcastd 104(%rax), %ymm19
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm5, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm6, %zmm24, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} ymm6 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm14
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm12, %xmm30
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermd %ymm23, %ymm17, %ymm17
+; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm19
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm17, %zmm20
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm8, %zmm18, %zmm20
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm1, %zmm27, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm7, %zmm1, %zmm5
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm15[4],xmm9[5],xmm15[5],xmm9[6],xmm15[6],xmm9[7],xmm15[7]
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm15[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm8
; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm24 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm13
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0],xmm0[1],xmm9[2,3],xmm0[4],xmm9[5,6],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm25, %zmm0
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm3, %ymm30, %ymm7
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm22, %xmm13
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm10, %xmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm28, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} ymm2 = ymm6[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm26 = ymm2[3,3,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm29 = ymm2[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm1
-; AVX512DQ-FCP-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vpbroadcastd 104(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm22, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm15
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm9 = xmm6[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm16, %zmm10
-; AVX512DQ-FCP-NEXT: vpermd %ymm14, %ymm31, %ymm9
-; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm18
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm9, %zmm9
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm10, %zmm30, %zmm9
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm2
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm17, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm12, %xmm0
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm8[0],xmm0[1],xmm8[2,3],xmm0[4],xmm8[5,6],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm25, %zmm0
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm4
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm28, %zmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm0[4],xmm9[4],xmm0[5],xmm9[5],xmm0[6],xmm9[6],xmm0[7],xmm9[7]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm0[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm12[0],xmm7[1],xmm12[2,3],xmm7[4],xmm12[5,6],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm0, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm8, %zmm28, %zmm7
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm3[4],xmm11[5],xmm3[5],xmm11[6],xmm3[6],xmm11[7],xmm3[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm3
-; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm4, %zmm22, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm7, %ymm0, %ymm29
+; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm24, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm7, %zmm1, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm11, %ymm0, %ymm26
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm1
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm29[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm12
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd (%rsp), %xmm1, %xmm1 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm0, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm13, %ymm1, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm16, %ymm2, %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm2 # 64-byte Folded Reload
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd (%rsp), %xmm4, %xmm4 # 16-byte Folded Reload
; AVX512DQ-FCP-NEXT: # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm5 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm7 = mem[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4],ymm6[5],ymm7[6,7,8,9],ymm6[10],ymm7[11,12],ymm6[13],ymm7[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm8 = mem[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4],ymm8[5],ymm7[6,7,8,9],ymm8[10],ymm7[11,12],ymm8[13],ymm7[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpandn %ymm8, %ymm11, %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512DQ-FCP-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm10 = mem[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7,8,9],ymm10[10],ymm9[11,12],ymm10[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm13
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm10, %zmm25, %zmm1
+; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm10, %ymm10
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm11[16,17],zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm10
+; AVX512DQ-FCP-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm11 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm14
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
; 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), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm4, %zmm25, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm5 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm11, %zmm28, %zmm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm28, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm10 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm10
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, 448(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 704(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 640(%rax)
-; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 576(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 512(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 832(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 768(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 448(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 704(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 640(%rax)
+; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 576(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 512(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 832(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 768(%rax)
; AVX512DQ-FCP-NEXT: addq $1496, %rsp # imm = 0x5D8
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-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 f56c43eb49df2..06d390f053c7e 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
@@ -4879,12 +4879,10 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
; AVX512-NEXT: vpternlogq $248, %zmm28, %zmm7, %zmm2
-; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm26[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vporq %zmm5, %zmm7, %zmm5
-; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm25[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm9 = zmm27[2,2,3,3,6,6,7,7]
-; AVX512-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512-NEXT: vporq %zmm24, %zmm26, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512-NEXT: vporq %zmm25, %zmm27, %zmm7
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,2,3,3,6,6,7,7]
; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm7
; AVX512-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm30
@@ -4918,165 +4916,165 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP: # %bb.0:
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm3
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm3, %ymm17
; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm15[8,u],zero,xmm15[7],zero,xmm15[9,u,11,u],zero,xmm15[10],zero,xmm15[12,u],zero
-; AVX512-FCP-NEXT: vporq %xmm2, %xmm4, %xmm18
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm5, %xmm4
+; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm28
+; AVX512-FCP-NEXT: vporq %xmm1, %xmm4, %xmm18
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm1
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm4
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm6
; AVX512-FCP-NEXT: vporq %ymm4, %ymm6, %ymm19
; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm7
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %xmm13, %xmm29
; AVX512-FCP-NEXT: vporq %xmm4, %xmm8, %xmm20
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm8, %ymm31
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm21
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
-; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm22
+; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm21
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm8
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm30
+; AVX512-FCP-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-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm22
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
-; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm24
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm8
+; AVX512-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm15
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm15, %zmm24
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512-FCP-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-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm11
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm26
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm10
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm10
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm10, %ymm23
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm12
+; AVX512-FCP-NEXT: vporq %ymm10, %ymm12, %ymm23
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm10
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm10, %ymm25
; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm10
; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm10, %ymm0
; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm27
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm9
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm9, %ymm27
; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9
; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm14[8,u],zero,xmm14[7],zero,xmm14[9,u,11,u],zero,xmm14[10],zero,xmm14[12,u],zero
-; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm3
+; AVX512-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm14, %xmm2
+; AVX512-FCP-NEXT: vporq %xmm0, %xmm2, %xmm28
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm1
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm5
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm5
-; AVX512-FCP-NEXT: vpor %xmm1, %xmm5, %xmm5
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2
-; AVX512-FCP-NEXT: vpor %ymm1, %ymm2, %ymm11
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[26],zero,ymm12[28],zero,zero,ymm12[27],zero,ymm12[29],zero,ymm12[31],zero,zero,ymm12[30],zero
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm13[6],zero,xmm13[8,u],zero,xmm13[7],zero,xmm13[9],zero,xmm13[11,u],zero,xmm13[10],zero,xmm13[12]
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm2
+; AVX512-FCP-NEXT: vporq %xmm0, %xmm2, %xmm29
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm2
; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm8
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm4
; AVX512-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm2
; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
-; AVX512-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm10
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm10, %ymm10
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512-FCP-NEXT: vpandn %ymm10, %ymm11, %ymm10
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm2
; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm10
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512-FCP-NEXT: vpermd %ymm10, %ymm28, %ymm30
-; AVX512-FCP-NEXT: vpandnq %ymm30, %ymm29, %ymm30
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm4
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm12 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512-FCP-NEXT: vpermd %ymm10, %ymm12, %ymm31
+; AVX512-FCP-NEXT: vpandnq %ymm31, %ymm30, %ymm31
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm10, %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm8, %zmm8
; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm9, %zmm3
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3],xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm5
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm10
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm28, %zmm9, %zmm9
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3],xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm13, %xmm13
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm29, %zmm13, %zmm13
+; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm28
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm28, %zmm10
; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm11
-; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm6
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3],xmm6[4],xmm15[4],xmm6[5],xmm15[5],xmm6[6],xmm15[6],xmm6[7],xmm15[7]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm1
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm21[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vporq %zmm6, %zmm7, %zmm6
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm26[2,2,3,3,6,6,7,7]
-; AVX512-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm9, %zmm7
-; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm8
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm8
-; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm18[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm20[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm7, %zmm7
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
-; AVX512-FCP-NEXT: vpternlogq $248, %zmm12, %zmm7, %zmm2
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm25[2,2,3,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm23, %zmm6
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm16[2,2,3,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm27, %zmm7
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm11, %zmm9, %zmm1
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512-FCP-NEXT: vpermd %zmm10, %zmm3, %zmm3
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512-FCP-NEXT: vpermd %zmm5, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 64(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm0
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm1
+; AVX512-FCP-NEXT: vporq %zmm21, %zmm22, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vporq %zmm24, %zmm26, %zmm5
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm6, %zmm5
+; AVX512-FCP-NEXT: vpermt2d %zmm28, %zmm12, %zmm4
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm5, %zmm30, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 256(%r9)
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm20[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm5, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $248, %zmm11, %zmm4, %zmm2
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm25[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm23, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm16[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm27, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm5, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm1
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm8
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm9[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm13[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
+; AVX512-FCP-NEXT: vpermd %zmm10, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512-FCP-NEXT: vpermd %zmm28, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm3
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 128(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%r9)
; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
@@ -5224,12 +5222,10 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm7
; AVX512DQ-NEXT: vpternlogq $248, %zmm28, %zmm7, %zmm2
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm26[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vporq %zmm5, %zmm7, %zmm5
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm25[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm9 = zmm27[2,2,3,3,6,6,7,7]
-; AVX512DQ-NEXT: vporq %zmm7, %zmm9, %zmm7
+; AVX512DQ-NEXT: vporq %zmm24, %zmm26, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512DQ-NEXT: vporq %zmm25, %zmm27, %zmm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,2,3,3,6,6,7,7]
; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm7
; AVX512DQ-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm30
@@ -5263,165 +5259,165 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP: # %bb.0:
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm3
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm17
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm3, %ymm17
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm31
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm15[8,u],zero,xmm15[7],zero,xmm15[9,u,11,u],zero,xmm15[10],zero,xmm15[12,u],zero
-; AVX512DQ-FCP-NEXT: vporq %xmm2, %xmm4, %xmm18
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[8],zero,xmm3[u,7],zero,xmm3[9],zero,xmm3[u],zero,xmm3[u,10],zero,xmm3[12],zero,xmm3[u,11]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm5, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm28
+; AVX512DQ-FCP-NEXT: vporq %xmm1, %xmm4, %xmm18
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm1
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm11
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm6
; AVX512DQ-FCP-NEXT: vporq %ymm4, %ymm6, %ymm19
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm6[6],zero,xmm6[8,u],zero,xmm6[7],zero,xmm6[9],zero,xmm6[11,u],zero,xmm6[10],zero,xmm6[12]
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm13, %xmm29
; AVX512DQ-FCP-NEXT: vporq %xmm4, %xmm8, %xmm20
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[26],zero,ymm1[28],zero,zero,ymm1[27],zero,ymm1[29],zero,ymm1[31],zero,zero,ymm1[30],zero
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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-FCP-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm8, %ymm31
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm21
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
-; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm22
+; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm21
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm30
+; AVX512DQ-FCP-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-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm22
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
-; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm24
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm8
+; AVX512DQ-FCP-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-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm15
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm15, %zmm24
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512DQ-FCP-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-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm11
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm26
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm10
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm10
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm10, %ymm23
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm25
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm12
+; AVX512DQ-FCP-NEXT: vporq %ymm10, %ymm12, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm11, %ymm10
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm10, %ymm25
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm10
; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm10, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm27
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm12, %ymm9
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm9, %ymm27
; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[8],zero,xmm9[u,7],zero,xmm9[9],zero,xmm9[u],zero,xmm9[u,10],zero,xmm9[12],zero,xmm9[u,11]
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm14[8,u],zero,xmm14[7],zero,xmm14[9,u,11,u],zero,xmm14[10],zero,xmm14[12,u],zero
-; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm14, %xmm2
+; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm2, %xmm28
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm5
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm5
-; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm2, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[26],zero,ymm12[28],zero,zero,ymm12[27],zero,ymm12[29],zero,ymm12[31],zero,zero,ymm12[30],zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm13[6],zero,xmm13[8,u],zero,xmm13[7],zero,xmm13[9],zero,xmm13[11,u],zero,xmm13[10],zero,xmm13[12]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm2
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm2
+; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm2, %xmm29
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm2
; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm4
; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm2
; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [1,1,2,2,2,2,2,2]
-; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm10, %ymm10
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512DQ-FCP-NEXT: vpandn %ymm10, %ymm12, %ymm10
+; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm10, %ymm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512DQ-FCP-NEXT: vpandn %ymm10, %ymm11, %ymm10
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm10
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm28 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm29 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm28, %ymm30
-; AVX512DQ-FCP-NEXT: vpandnq %ymm30, %ymm29, %ymm30
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm4
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm12 = [4,6,5,5,5,5,4,6,30,30,30,30,31,31,31,31]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm12, %ymm31
+; AVX512DQ-FCP-NEXT: vpandnq %ymm31, %ymm30, %ymm31
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm10, %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm8, %zmm8
; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm14[0],xmm9[1],xmm14[1],xmm9[2],xmm14[2],xmm9[3],xmm14[3],xmm9[4],xmm14[4],xmm9[5],xmm14[5],xmm9[6],xmm14[6],xmm9[7],xmm14[7]
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm3, %zmm9, %zmm3
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3],xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm5
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm10
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm28, %zmm9, %zmm9
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3],xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm29, %zmm13, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm28
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm28, %zmm10
; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [2,2,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm6
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3],xmm6[4],xmm15[4],xmm6[5],xmm15[5],xmm6[6],xmm15[6],xmm6[7],xmm15[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm21[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm22[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm6, %zmm7, %zmm6
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm26[2,2,3,3,6,6,7,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm9, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm9, %zmm7
-; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm28, %zmm8
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm7, %zmm29, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm18[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm20[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm7, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm12, %zmm7, %zmm2
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm25[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm23, %zmm6
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm16[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm27, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm6, %zmm8, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm11, %zmm9, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm10, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 64(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm0
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm1
+; AVX512DQ-FCP-NEXT: vporq %zmm21, %zmm22, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm24, %zmm26, %zmm5
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm6, %zmm5
+; AVX512DQ-FCP-NEXT: vpermt2d %zmm28, %zmm12, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm5, %zmm30, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 256(%r9)
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm18[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm20[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $248, %zmm11, %zmm4, %zmm2
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm25[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm23, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm16[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm27, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm5, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm8
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm9[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm13[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,0,0,1,1,9,9,0,10,10,10,10,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm10, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512DQ-FCP-NEXT: vpermd %zmm28, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 128(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 64(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%r9)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll
index e43aa56c96c28..60af864597f4f 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
@@ -5405,304 +5405,289 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-LABEL: store_i8_stride6_vf64:
; AVX512: # %bb.0:
-; AVX512-NEXT: subq $328, %rsp # imm = 0x148
-; AVX512-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512-NEXT: vmovdqa (%rdi), %xmm10
-; AVX512-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512-NEXT: vmovdqa (%r9), %xmm11
-; AVX512-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512-NEXT: vmovdqa 32(%rcx), %ymm8
-; AVX512-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512-NEXT: vmovdqa 32(%rsi), %ymm6
-; AVX512-NEXT: vmovdqa 32(%rdi), %ymm12
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm14 = [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]
-; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm4 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512-NEXT: vpshufb %xmm4, %xmm2, %xmm0
-; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm1
+; AVX512-NEXT: subq $456, %rsp # imm = 0x1C8
+; AVX512-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm3
+; AVX512-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512-NEXT: vmovdqa (%r8), %xmm10
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512-NEXT: vmovdqa (%r9), %xmm12
+; AVX512-NEXT: vmovdqa 32(%r9), %xmm8
+; AVX512-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512-NEXT: vpshufb %xmm9, %xmm3, %xmm0
+; AVX512-NEXT: vpshufb %xmm9, %xmm2, %xmm1
; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm16
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm27
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512-NEXT: vmovdqa64 %xmm2, %xmm31
; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5]
; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
-; AVX512-NEXT: vmovdqa %xmm7, %xmm3
-; AVX512-NEXT: vpshufb %xmm0, %xmm7, %xmm2
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
-; AVX512-NEXT: vpshufb %xmm1, %xmm7, %xmm7
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm26
-; AVX512-NEXT: vinserti32x4 $2, %xmm2, %zmm7, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %xmm4, %xmm5, %xmm2
-; AVX512-NEXT: vpshufb %xmm4, %xmm10, %xmm4
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3],xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
-; AVX512-NEXT: vmovdqa64 %xmm10, %xmm29
-; AVX512-NEXT: vmovdqa64 %xmm5, %xmm30
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,2,1,4,5,6,7]
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,6,5]
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
+; AVX512-NEXT: vpshufb %xmm0, %xmm4, %xmm2
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
+; AVX512-NEXT: vpshufb %xmm1, %xmm4, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2
; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %xmm0, %xmm11, %xmm0
-; AVX512-NEXT: vpshufb %xmm1, %xmm11, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm11, %xmm31
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
+; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm4
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
+; AVX512-NEXT: vpshufb %xmm3, %xmm8, %xmm5
+; AVX512-NEXT: vmovdqa64 %xmm8, %xmm27
+; AVX512-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rdx), %ymm8
+; AVX512-NEXT: vpshufb %xmm9, %xmm6, %xmm4
+; AVX512-NEXT: vpshufb %xmm9, %xmm7, %xmm5
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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]
+; AVX512-NEXT: vmovdqa64 %xmm7, %xmm18
+; AVX512-NEXT: vmovdqa64 %xmm6, %xmm20
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,3,2,1,4,5,6,7]
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,6,5]
+; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%rsi), %ymm4
+; AVX512-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512-NEXT: vpshufb %xmm0, %xmm10, %xmm0
+; AVX512-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm10, %xmm22
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpshufb %xmm2, %xmm12, %xmm0
+; AVX512-NEXT: vpshufb %xmm3, %xmm12, %xmm1
+; AVX512-NEXT: vmovdqa64 %xmm12, %xmm21
; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512-NEXT: vpshufb %ymm7, %ymm6, %ymm0
-; AVX512-NEXT: vpshufb %ymm7, %ymm12, %ymm1
+; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm15 = [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]
+; AVX512-NEXT: vpshufb %ymm15, %ymm4, %ymm0
+; AVX512-NEXT: vpshufb %ymm15, %ymm11, %ymm1
; AVX512-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]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm12[8],ymm6[8],ymm12[9],ymm6[9],ymm12[10],ymm6[10],ymm12[11],ymm6[11],ymm12[12],ymm6[12],ymm12[13],ymm6[13],ymm12[14],ymm6[14],ymm12[15],ymm6[15],ymm12[24],ymm6[24],ymm12[25],ymm6[25],ymm12[26],ymm6[26],ymm12[27],ymm6[27],ymm12[28],ymm6[28],ymm12[29],ymm6[29],ymm12[30],ymm6[30],ymm12[31],ymm6[31]
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
-; AVX512-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm13, %ymm1, %ymm1
+; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm4[8],ymm11[9],ymm4[9],ymm11[10],ymm4[10],ymm11[11],ymm4[11],ymm11[12],ymm4[12],ymm11[13],ymm4[13],ymm11[14],ymm4[14],ymm11[15],ymm4[15],ymm11[24],ymm4[24],ymm11[25],ymm4[25],ymm11[26],ymm4[26],ymm11[27],ymm4[27],ymm11[28],ymm4[28],ymm11[29],ymm4[29],ymm11[30],ymm4[30],ymm11[31],ymm4[31]
+; AVX512-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
+; AVX512-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm6, %ymm1, %ymm1
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512-NEXT: vpshufb %ymm14, %ymm8, %ymm0
-; AVX512-NEXT: vpshufb %ymm14, %ymm9, %ymm1
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm1 = 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]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15],ymm9[24],ymm8[24],ymm9[25],ymm8[25],ymm9[26],ymm8[26],ymm9[27],ymm8[27],ymm9[28],ymm8[28],ymm9[29],ymm8[29],ymm9[30],ymm8[30],ymm9[31],ymm8[31]
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
-; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm12 = [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]
+; AVX512-NEXT: vpshufb %ymm12, %ymm13, %ymm0
+; AVX512-NEXT: vpshufb %ymm12, %ymm8, %ymm1
+; AVX512-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]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm8[8],ymm13[8],ymm8[9],ymm13[9],ymm8[10],ymm13[10],ymm8[11],ymm13[11],ymm8[12],ymm13[12],ymm8[13],ymm13[13],ymm8[14],ymm13[14],ymm8[15],ymm13[15],ymm8[24],ymm13[24],ymm8[25],ymm13[25],ymm8[26],ymm13[26],ymm8[27],ymm13[27],ymm8[28],ymm13[28],ymm8[29],ymm13[29],ymm8[30],ymm13[30],ymm8[31],ymm13[31]
+; AVX512-NEXT: vmovdqa64 %ymm13, %ymm17
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
+; AVX512-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm7, %ymm1, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r8), %ymm13
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
+; AVX512-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm5, %ymm13, %ymm0
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
+; AVX512-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm4, %ymm13, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa 32(%r9), %ymm14
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm0
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm4, %ymm10
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
-; AVX512-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm11
-; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm3
-; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rsi), %ymm15
-; AVX512-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512-NEXT: vpshufb %ymm7, %ymm15, %ymm11
-; AVX512-NEXT: vpshufb %ymm7, %ymm10, %ymm7
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm11[0],ymm7[1],ymm11[1],ymm7[2],ymm11[2],ymm7[3],ymm11[3],ymm7[4],ymm11[4],ymm7[5],ymm11[5],ymm7[6],ymm11[6],ymm7[7],ymm11[7],ymm7[16],ymm11[16],ymm7[17],ymm11[17],ymm7[18],ymm11[18],ymm7[19],ymm11[19],ymm7[20],ymm11[20],ymm7[21],ymm11[21],ymm7[22],ymm11[22],ymm7[23],ymm11[23]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm10[8],ymm15[8],ymm10[9],ymm15[9],ymm10[10],ymm15[10],ymm10[11],ymm15[11],ymm10[12],ymm15[12],ymm10[13],ymm15[13],ymm10[14],ymm15[14],ymm10[15],ymm15[15],ymm10[24],ymm15[24],ymm10[25],ymm15[25],ymm10[26],ymm15[26],ymm10[27],ymm15[27],ymm10[28],ymm15[28],ymm10[29],ymm15[29],ymm10[30],ymm15[30],ymm10[31],ymm15[31]
-; AVX512-NEXT: vpshufb %ymm13, %ymm11, %ymm11
-; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm25
-; AVX512-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512-NEXT: vpshufb %ymm14, %ymm7, %ymm13
-; AVX512-NEXT: vpshufb %ymm14, %ymm11, %ymm5
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm13[0],ymm5[1],ymm13[1],ymm5[2],ymm13[2],ymm5[3],ymm13[3],ymm5[4],ymm13[4],ymm5[5],ymm13[5],ymm5[6],ymm13[6],ymm5[7],ymm13[7],ymm5[16],ymm13[16],ymm5[17],ymm13[17],ymm5[18],ymm13[18],ymm5[19],ymm13[19],ymm5[20],ymm13[20],ymm5[21],ymm13[21],ymm5[22],ymm13[22],ymm5[23],ymm13[23]
-; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm13 = ymm11[8],ymm7[8],ymm11[9],ymm7[9],ymm11[10],ymm7[10],ymm11[11],ymm7[11],ymm11[12],ymm7[12],ymm11[13],ymm7[13],ymm11[14],ymm7[14],ymm11[15],ymm7[15],ymm11[24],ymm7[24],ymm11[25],ymm7[25],ymm11[26],ymm7[26],ymm11[27],ymm7[27],ymm11[28],ymm7[28],ymm11[29],ymm7[29],ymm11[30],ymm7[30],ymm11[31],ymm7[31]
-; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm24
-; AVX512-NEXT: vmovdqa (%r9), %ymm3
-; AVX512-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm23
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[2],ymm7[2],ymm11[3],ymm7[3],ymm11[4],ymm7[4],ymm11[5],ymm7[5],ymm11[6],ymm7[6],ymm11[7],ymm7[7],ymm11[16],ymm7[16],ymm11[17],ymm7[17],ymm11[18],ymm7[18],ymm11[19],ymm7[19],ymm11[20],ymm7[20],ymm11[21],ymm7[21],ymm11[22],ymm7[22],ymm11[23],ymm7[23]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512-NEXT: vmovdqa64 %xmm29, %xmm0
-; AVX512-NEXT: vmovdqa64 %xmm30, %xmm1
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[2],ymm15[2],ymm10[3],ymm15[3],ymm10[4],ymm15[4],ymm10[5],ymm15[5],ymm10[6],ymm15[6],ymm10[7],ymm15[7],ymm10[16],ymm15[16],ymm10[17],ymm15[17],ymm10[18],ymm15[18],ymm10[19],ymm15[19],ymm10[20],ymm15[20],ymm10[21],ymm15[21],ymm10[22],ymm15[22],ymm10[23],ymm15[23]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm10[8],xmm2[8],xmm10[9],xmm2[9],xmm10[10],xmm2[10],xmm10[11],xmm2[11],xmm10[12],xmm2[12],xmm10[13],xmm2[13],xmm10[14],xmm2[14],xmm10[15],xmm2[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm15 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512-NEXT: vpshufb %xmm15, %xmm5, %xmm0
+; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm28
+; AVX512-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512-NEXT: vpshufb %ymm15, %ymm1, %ymm10
+; AVX512-NEXT: vpshufb %ymm15, %ymm0, %ymm15
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
+; AVX512-NEXT: vpshufb %ymm6, %ymm15, %ymm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm26
+; AVX512-NEXT: vmovdqa (%rcx), %ymm10
+; AVX512-NEXT: vmovdqa (%rdx), %ymm15
+; AVX512-NEXT: vpshufb %ymm12, %ymm10, %ymm6
+; AVX512-NEXT: vpshufb %ymm12, %ymm15, %ymm9
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[4],ymm6[4],ymm9[5],ymm6[5],ymm9[6],ymm6[6],ymm9[7],ymm6[7],ymm9[16],ymm6[16],ymm9[17],ymm6[17],ymm9[18],ymm6[18],ymm9[19],ymm6[19],ymm9[20],ymm6[20],ymm9[21],ymm6[21],ymm9[22],ymm6[22],ymm9[23],ymm6[23]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm15[8],ymm10[8],ymm15[9],ymm10[9],ymm15[10],ymm10[10],ymm15[11],ymm10[11],ymm15[12],ymm10[12],ymm15[13],ymm10[13],ymm15[14],ymm10[14],ymm15[15],ymm10[15],ymm15[24],ymm10[24],ymm15[25],ymm10[25],ymm15[26],ymm10[26],ymm15[27],ymm10[27],ymm15[28],ymm10[28],ymm15[29],ymm10[29],ymm15[30],ymm10[30],ymm15[31],ymm10[31]
+; AVX512-NEXT: vpshufb %ymm7, %ymm9, %ymm9
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm24
+; AVX512-NEXT: vmovdqa (%r8), %ymm6
+; AVX512-NEXT: vpshufb %ymm5, %ymm6, %ymm5
+; AVX512-NEXT: vpshufb %ymm4, %ymm6, %ymm4
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm23
+; AVX512-NEXT: vmovdqa (%r9), %ymm4
+; AVX512-NEXT: vpshufb %ymm3, %ymm4, %ymm3
+; AVX512-NEXT: vpshufb %ymm2, %ymm4, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm25
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23]
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512-NEXT: vmovdqa64 %xmm20, %xmm3
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; AVX512-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512-NEXT: vmovdqa (%rcx), %xmm15
+; AVX512-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm15[8],xmm2[9],xmm15[9],xmm2[10],xmm15[10],xmm2[11],xmm15[11],xmm2[12],xmm15[12],xmm2[13],xmm15[13],xmm2[14],xmm15[14],xmm2[15],xmm15[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512-NEXT: vpshufb %xmm3, %xmm1, %xmm10
; AVX512-NEXT: vmovdqa 32(%rcx), %xmm1
; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23]
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm30
-; AVX512-NEXT: vmovdqa64 %xmm16, %xmm5
-; AVX512-NEXT: vmovdqa64 %xmm27, %xmm8
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm5[8],xmm8[8],xmm5[9],xmm8[9],xmm5[10],xmm8[10],xmm5[11],xmm8[11],xmm5[12],xmm8[12],xmm5[13],xmm8[13],xmm5[14],xmm8[14],xmm5[15],xmm8[15]
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm9 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512-NEXT: vpshufb %xmm9, %xmm7, %xmm11
-; AVX512-NEXT: vpshufb %xmm9, %xmm8, %xmm5
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm21
-; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm12[0],ymm6[0],ymm12[1],ymm6[1],ymm12[2],ymm6[2],ymm12[3],ymm6[3],ymm12[4],ymm6[4],ymm12[5],ymm6[5],ymm12[6],ymm6[6],ymm12[7],ymm6[7],ymm12[16],ymm6[16],ymm12[17],ymm6[17],ymm12[18],ymm6[18],ymm12[19],ymm6[19],ymm12[20],ymm6[20],ymm12[21],ymm6[21],ymm12[22],ymm6[22],ymm12[23],ymm6[23]
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm5
-; AVX512-NEXT: vpshufb %xmm6, %xmm5, %xmm13
-; AVX512-NEXT: vmovdqa64 %xmm26, %xmm5
-; AVX512-NEXT: vpshufb %xmm6, %xmm5, %xmm5
-; AVX512-NEXT: vmovdqa64 %ymm5, %ymm31
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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]
-; AVX512-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm8, %ymm3, %ymm9
-; AVX512-NEXT: vpshufb %ymm8, %ymm4, %ymm3
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
+; AVX512-NEXT: vpshufb %xmm3, %xmm9, %xmm3
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm20
+; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[1],ymm3[1],ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[4],ymm3[4],ymm8[5],ymm3[5],ymm8[6],ymm3[6],ymm8[7],ymm3[7],ymm8[16],ymm3[16],ymm8[17],ymm3[17],ymm8[18],ymm3[18],ymm8[19],ymm3[19],ymm8[20],ymm3[20],ymm8[21],ymm3[21],ymm8[22],ymm3[22],ymm8[23],ymm3[23]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm17
+; AVX512-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm31, %xmm7
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm3[8],xmm7[9],xmm3[9],xmm7[10],xmm3[10],xmm7[11],xmm3[11],xmm7[12],xmm3[12],xmm7[13],xmm3[13],xmm7[14],xmm3[14],xmm7[15],xmm3[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512-NEXT: vpshufb %xmm8, %xmm5, %xmm9
+; AVX512-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm11[0],ymm3[0],ymm11[1],ymm3[1],ymm11[2],ymm3[2],ymm11[3],ymm3[3],ymm11[4],ymm3[4],ymm11[5],ymm3[5],ymm11[6],ymm3[6],ymm11[7],ymm3[7],ymm11[16],ymm3[16],ymm11[17],ymm3[17],ymm11[18],ymm3[18],ymm11[19],ymm3[19],ymm11[20],ymm3[20],ymm11[21],ymm3[21],ymm11[22],ymm3[22],ymm11[23],ymm3[23]
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
+; AVX512-NEXT: vmovdqa64 %xmm22, %xmm5
+; AVX512-NEXT: vpshufb %xmm3, %xmm5, %xmm7
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm5
+; AVX512-NEXT: vpshufb %xmm3, %xmm5, %xmm3
; AVX512-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512-NEXT: vpshufb %xmm14, %xmm1, %xmm3
-; AVX512-NEXT: vpshufb %xmm14, %xmm0, %xmm4
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [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]
+; AVX512-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm11, %ymm6, %ymm5
+; AVX512-NEXT: vpshufb %ymm11, %ymm13, %ymm6
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
+; AVX512-NEXT: vmovdqa64 %xmm21, %xmm3
+; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm13
+; AVX512-NEXT: vmovdqa64 %xmm27, %xmm3
+; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm11
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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]
+; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm3, %ymm4, %ymm4
+; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512-NEXT: vpshufb %xmm12, %xmm1, %xmm14
+; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm8
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm27
-; AVX512-NEXT: vpshufb %xmm14, %xmm2, %xmm0
-; AVX512-NEXT: vpshufb %xmm14, %xmm10, %xmm1
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512-NEXT: vmovdqa (%r8), %xmm1
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3],xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512-NEXT: vpshufb %xmm12, %xmm15, %xmm1
+; AVX512-NEXT: vpshufb %xmm12, %xmm2, %xmm8
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm4 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
-; AVX512-NEXT: vpshufb %xmm4, %xmm1, %xmm6
-; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm15
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
-; AVX512-NEXT: vpshufb %xmm2, %xmm3, %xmm10
-; AVX512-NEXT: vmovdqa64 %ymm10, %ymm17
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm4 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
-; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm10
-; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm14
-; AVX512-NEXT: vpshufb %xmm4, %xmm1, %xmm0
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512-NEXT: vmovdqa (%r8), %ymm1
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
-; AVX512-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm2, %ymm1, %ymm5
-; AVX512-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
-; AVX512-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm7
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
-; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm12
-; AVX512-NEXT: vpshufb %ymm4, %ymm1, %ymm3
-; AVX512-NEXT: vpshufb %ymm0, %ymm1, %ymm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm18[0,0,0,1]
-; AVX512-NEXT: vprold $16, %ymm28, %ymm1
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm8
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
-; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
+; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm10[0,0,0,1]
+; AVX512-NEXT: vprold $16, %ymm16, %ymm8
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
+; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm6[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm13[0,0,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm20[0,0,0,1]
+; AVX512-NEXT: vprold $16, %ymm17, %ymm14
+; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm31[0,0,0,1]
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm1
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm1
-; AVX512-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512-NEXT: vpternlogq $184, %ymm0, %ymm6, %ymm5
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512-NEXT: vpternlogq $184, %ymm1, %ymm0, %ymm28
-; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm1
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm28[0,1,2,3],zmm1[4,5,6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm13, %zmm1
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
-; AVX512-NEXT: vpternlogd $184, %zmm5, %zmm8, %zmm1
-; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm29[0,0,0,1]
-; AVX512-NEXT: vprold $16, %ymm30, %ymm9
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm21[0,0,0,1]
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm13
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
-; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm28 = ymm31[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm29 = ymm22[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm30 = ymm17[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm16[0,0,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm22[0,0,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
-; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm9
-; AVX512-NEXT: vpternlogq $226, %zmm5, %zmm6, %zmm9
-; AVX512-NEXT: vextracti64x4 $1, %zmm9, %ymm5
-; AVX512-NEXT: vpternlogq $184, %ymm5, %ymm6, %ymm2
-; AVX512-NEXT: vpternlogq $184, %ymm9, %ymm0, %ymm15
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm2
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm15[0,1,2,3],zmm2[4,5,6,7]
-; AVX512-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm5
-; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm8, %zmm5
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm27[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm8
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm9, %zmm8
-; AVX512-NEXT: vextracti64x4 $1, %zmm8, %ymm2
-; AVX512-NEXT: vpternlogq $184, %ymm2, %ymm9, %ymm30
-; AVX512-NEXT: vpternlogq $184, %ymm8, %ymm6, %ymm10
-; AVX512-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm2
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm10[0,1,2,3],zmm2[4,5,6,7]
-; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
-; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm10, %zmm8
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm26[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm11 = mem[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm9, %zmm11
-; AVX512-NEXT: vextracti64x4 $1, %zmm11, %ymm2
-; AVX512-NEXT: vpternlogq $184, %ymm2, %ymm9, %ymm14
-; AVX512-NEXT: vpternlogq $184, %ymm11, %ymm6, %ymm31
-; AVX512-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm2
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm31[0,1,2,3],zmm2[4,5,6,7]
-; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm11 = mem[0,0,0,1,4,4,4,5]
-; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm10, %zmm11
-; AVX512-NEXT: vpermq $234, (%rsp), %zmm2 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm10 = mem[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm10
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm25[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm13 = zmm24[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm13
-; AVX512-NEXT: vextracti64x4 $1, %zmm10, %ymm2
-; AVX512-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm7
-; AVX512-NEXT: vextracti64x4 $1, %zmm13, %ymm2
-; AVX512-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm3
-; AVX512-NEXT: vpternlogq $184, %ymm10, %ymm9, %ymm12
-; AVX512-NEXT: vpternlogq $184, %ymm13, %ymm9, %ymm4
-; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm12[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
-; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm3 = zmm23[2,2,2,3,6,6,6,7]
-; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm2
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512-NEXT: vpternlogd $184, %zmm8, %zmm5, %zmm2
+; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm14[2,2,2,3]
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7
+; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm1
+; AVX512-NEXT: vpternlogq $226, %zmm7, %zmm9, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6
+; AVX512-NEXT: vpternlogd $184, %zmm1, %zmm5, %zmm6
+; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm12, %zmm1
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
+; AVX512-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm2
+; AVX512-NEXT: vpternlogd $184, %zmm6, %zmm4, %zmm2
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm4, %zmm3
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpternlogq $184, %zmm3, %zmm9, %zmm0
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
+; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm4, %zmm6
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpternlogq $184, %zmm6, %zmm9, %zmm0
+; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5]
+; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm6
+; AVX512-NEXT: vpermq $234, (%rsp), %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm7
+; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpternlogq $184, %zmm5, %zmm4, %zmm0
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm23[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpternlogq $184, %zmm7, %zmm4, %zmm5
+; AVX512-NEXT: vpermq {{.*#+}} zmm4 = zmm28[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm7 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
+; AVX512-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm4
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm25[2,2,2,3,6,6,6,7]
+; AVX512-NEXT: vpternlogd $184, %zmm5, %zmm7, %zmm0
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm11, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm8, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm0, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm4, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm6, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax)
; AVX512-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512-NEXT: addq $328, %rsp # imm = 0x148
+; AVX512-NEXT: addq $456, %rsp # imm = 0x1C8
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i8_stride6_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $392, %rsp # imm = 0x188
+; AVX512-FCP-NEXT: subq $424, %rsp # imm = 0x1A8
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm4
; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-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]
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm1
; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm2
; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm2 = 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]
; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15],ymm3[24],ymm4[24],ymm3[25],ymm4[25],ymm3[26],ymm4[26],ymm3[27],ymm4[27],ymm3[28],ymm4[28],ymm3[29],ymm4[29],ymm3[30],ymm4[30],ymm3[31],ymm4[31]
@@ -5712,561 +5697,526 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm5
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm14 = [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]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm3
+; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm10 = [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]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm2
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm4
+; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm3
; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[16],ymm2[16],ymm3[17],ymm2[17],ymm3[18],ymm2[18],ymm3[19],ymm2[19],ymm3[20],ymm2[20],ymm3[21],ymm2[21],ymm3[22],ymm2[22],ymm3[23],ymm2[23]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15],ymm6[24],ymm5[24],ymm6[25],ymm5[25],ymm6[26],ymm5[26],ymm6[27],ymm5[27],ymm6[28],ymm5[28],ymm6[29],ymm5[29],ymm6[30],ymm5[30],ymm6[31],ymm5[31]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm25
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm30
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15],ymm4[24],ymm5[24],ymm4[25],ymm5[25],ymm4[26],ymm5[26],ymm4[27],ymm5[27],ymm4[28],ymm5[28],ymm4[29],ymm5[29],ymm4[30],ymm5[30],ymm4[31],ymm5[31]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm22
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm4
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm7
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm6
+; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm5
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm5
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm7, %ymm6
-; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm31
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm6
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm5
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[4],ymm5[4],ymm0[5],ymm5[5],ymm0[6],ymm5[6],ymm0[7],ymm5[7],ymm0[16],ymm5[16],ymm0[17],ymm5[17],ymm0[18],ymm5[18],ymm0[19],ymm5[19],ymm0[20],ymm5[20],ymm0[21],ymm5[21],ymm0[22],ymm5[22],ymm0[23],ymm5[23]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15],ymm7[24],ymm8[24],ymm7[25],ymm8[25],ymm7[26],ymm8[26],ymm7[27],ymm8[27],ymm7[28],ymm8[28],ymm7[29],ymm8[29],ymm7[30],ymm8[30],ymm7[31],ymm8[31]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
+; AVX512-FCP-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm6
+; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[1],ymm6[1],ymm0[2],ymm6[2],ymm0[3],ymm6[3],ymm0[4],ymm6[4],ymm0[5],ymm6[5],ymm0[6],ymm6[6],ymm0[7],ymm6[7],ymm0[16],ymm6[16],ymm0[17],ymm6[17],ymm0[18],ymm6[18],ymm0[19],ymm6[19],ymm0[20],ymm6[20],ymm0[21],ymm6[21],ymm0[22],ymm6[22],ymm0[23],ymm6[23]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15],ymm8[24],ymm7[24],ymm8[25],ymm7[25],ymm8[26],ymm7[26],ymm8[27],ymm7[27],ymm8[28],ymm7[28],ymm8[29],ymm7[29],ymm8[30],ymm7[30],ymm8[31],ymm7[31]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm8, %ymm18
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm1
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm11, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm12, %ymm1
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm15, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm11, %ymm1
; AVX512-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]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11],ymm12[12],ymm11[12],ymm12[13],ymm11[13],ymm12[14],ymm11[14],ymm12[15],ymm11[15],ymm12[24],ymm11[24],ymm12[25],ymm11[25],ymm12[26],ymm11[26],ymm12[27],ymm11[27],ymm12[28],ymm11[28],ymm12[29],ymm11[29],ymm12[30],ymm11[30],ymm12[31],ymm11[31]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm15[8],ymm11[9],ymm15[9],ymm11[10],ymm15[10],ymm11[11],ymm15[11],ymm11[12],ymm15[12],ymm11[13],ymm15[13],ymm11[14],ymm15[14],ymm11[15],ymm15[15],ymm11[24],ymm15[24],ymm11[25],ymm15[25],ymm11[26],ymm15[26],ymm11[27],ymm15[27],ymm11[28],ymm15[28],ymm11[29],ymm15[29],ymm11[30],ymm15[30],ymm11[31],ymm15[31]
; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm3
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm1
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm2
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm23
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm12, %xmm2
; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3],xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm18
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm4
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm2
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm4, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm19
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm26
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm4
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm27
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm15
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm15, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm15, %xmm1
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm28
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,2,2,3]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
-; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,2,2,3]
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm2, %xmm4
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm10
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3],xmm12[4],xmm0[4],xmm12[5],xmm0[5],xmm12[6],xmm0[6],xmm12[7],xmm0[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm16
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm26
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm13
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm13, %xmm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm13, %xmm3
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm29
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm14
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm14, %xmm4
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm4, %zmm21
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm4, %xmm5
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm5, %xmm5
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm25
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm2
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm2, %zmm27
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm3
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm3, %zmm28
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm7
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm3, %xmm9
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-FCP-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm30
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3],xmm10[4],xmm0[4],xmm10[5],xmm0[5],xmm10[6],xmm0[6],xmm10[7],xmm0[7]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX512-FCP-NEXT: vprold $16, %xmm10, %xmm10
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm10, %zmm23
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm4, %xmm0
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm10, %xmm14
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm14, %xmm14
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm29
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[4],ymm11[4],ymm12[5],ymm11[5],ymm12[6],ymm11[6],ymm12[7],ymm11[7],ymm12[16],ymm11[16],ymm12[17],ymm11[17],ymm12[18],ymm11[18],ymm12[19],ymm11[19],ymm12[20],ymm11[20],ymm12[21],ymm11[21],ymm12[22],ymm11[22],ymm12[23],ymm11[23]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm31
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm15[0],ymm11[1],ymm15[1],ymm11[2],ymm15[2],ymm11[3],ymm15[3],ymm11[4],ymm15[4],ymm11[5],ymm15[5],ymm11[6],ymm15[6],ymm11[7],ymm15[7],ymm11[16],ymm15[16],ymm11[17],ymm15[17],ymm11[18],ymm15[18],ymm11[19],ymm15[19],ymm11[20],ymm15[20],ymm11[21],ymm15[21],ymm11[22],ymm15[22],ymm11[23],ymm15[23]
; AVX512-FCP-NEXT: vprold $16, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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: vpunpcklbw {{.*#+}} ymm4 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[16],ymm8[16],ymm7[17],ymm8[17],ymm7[18],ymm8[18],ymm7[19],ymm8[19],ymm7[20],ymm8[20],ymm7[21],ymm8[21],ymm7[22],ymm8[22],ymm7[23],ymm8[23]
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
-; AVX512-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm4, %ymm4
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm11
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,0,0,1,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm11
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm10, %xmm3
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm16, %zmm11
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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]
-; AVX512-FCP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
-; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm15, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm5, %xmm17
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm5
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[4],ymm5[4],ymm0[5],ymm5[5],ymm0[6],ymm5[6],ymm0[7],ymm5[7],ymm0[16],ymm5[16],ymm0[17],ymm5[17],ymm0[18],ymm5[18],ymm0[19],ymm5[19],ymm0[20],ymm5[20],ymm0[21],ymm5[21],ymm0[22],ymm5[22],ymm0[23],ymm5[23]
-; AVX512-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]
-; AVX512-FCP-NEXT: vprold $16, %ymm0, %ymm0
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm1, %xmm2
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm2
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm4 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[16],mem[16],ymm0[17],mem[17],ymm0[18],mem[18],ymm0[19],mem[19],ymm0[20],mem[20],ymm0[21],mem[21],ymm0[22],mem[22],ymm0[23],mem[23]
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm5
-; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm4, %ymm4
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm14
-; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm8
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15]
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm15
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm8
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm8
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm15, %xmm4
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm8
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm9
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm2[0],ymm9[0],ymm2[1],ymm9[1],ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[4],ymm9[4],ymm2[5],ymm9[5],ymm2[6],ymm9[6],ymm2[7],ymm9[7],ymm2[16],ymm9[16],ymm2[17],ymm9[17],ymm2[18],ymm9[18],ymm2[19],ymm9[19],ymm2[20],ymm9[20],ymm2[21],ymm9[21],ymm2[22],ymm9[22],ymm2[23],ymm9[23]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
+; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm9, %ymm9
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm15
+; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm17
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,0,0,1,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm4, %zmm15
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm7, %zmm9, %zmm15
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm9
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm14, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm5
-; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm14, %xmm2
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm14
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm15, %xmm13
-; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm15, %ymm15
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm30 = ymm0[2,2,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm31 = ymm1[2,2,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm6[0,0,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[0,0,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm14[0,0,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm6
-; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm14
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm11, %ymm14, %ymm2
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm11, %ymm11
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm8, %ymm14, %ymm13
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm8
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm6
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm8, %ymm16, %ymm9
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm7
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm13[0,1,2,3],zmm7[4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm7, %zmm8, %zmm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm15, %zmm10, %zmm1
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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]
+; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm2
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm0
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm8
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[2],ymm8[2],ymm1[3],ymm8[3],ymm1[4],ymm8[4],ymm1[5],ymm8[5],ymm1[6],ymm8[6],ymm1[7],ymm8[7],ymm1[16],ymm8[16],ymm1[17],ymm8[17],ymm1[18],ymm8[18],ymm1[19],ymm8[19],ymm1[20],ymm8[20],ymm1[21],ymm8[21],ymm1[22],ymm8[22],ymm1[23],ymm8[23]
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
+; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vmovdqu (%rsp), %ymm5 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm5 = ymm5[0],mem[0],ymm5[1],mem[1],ymm5[2],mem[2],ymm5[3],mem[3],ymm5[4],mem[4],ymm5[5],mem[5],ymm5[6],mem[6],ymm5[7],mem[7],ymm5[16],mem[16],ymm5[17],mem[17],ymm5[18],mem[18],ymm5[19],mem[19],ymm5[20],mem[20],ymm5[21],mem[21],ymm5[22],mem[22],ymm5[23],mem[23]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm6
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm12[8],xmm6[8],xmm12[9],xmm6[9],xmm12[10],xmm6[10],xmm12[11],xmm6[11],xmm12[12],xmm6[12],xmm12[13],xmm6[13],xmm12[14],xmm6[14],xmm12[15],xmm6[15]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm6
+; AVX512-FCP-NEXT: vprold $16, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm3
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm9, %zmm6
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm13, %xmm3
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm3
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm6, %zmm10, %zmm3
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm5
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm5
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm2, %zmm5
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm11, %ymm16, %ymm12
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm6
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm8, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
-; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm3 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm3
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm2
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm2, %ymm14, %ymm22
-; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm6
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm6, %ymm2
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm2, %ymm14, %ymm30
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm1 = mem[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm1
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm24
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm3
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm24[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm7 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm8, %zmm7
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm6, %ymm2, %ymm31
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm3
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm31[0,1,2,3],zmm3[4,5,6,7]
-; AVX512-FCP-NEXT: vpermq $234, (%rsp), %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm3, %zmm8, %zmm6
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm23[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm8
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm3
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm4
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm29[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm27[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm9
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm9, %ymm3
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm8, %ymm16, %ymm5
-; AVX512-FCP-NEXT: vpternlogq $184, %ymm9, %ymm16, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm2
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm5[0,1,2,3],zmm2[4,5,6,7]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm0
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm1 = mem[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm1
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm4 = mem[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm4
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm23[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm4, %zmm2, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm24[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm4
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5]
; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm26[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm28[0,0,0,1,4,4,4,5]
-; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm4, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, (%rax)
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm31[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm25[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm29[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm3, %zmm9, %zmm0
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm27[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm5, %zmm9, %zmm2
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm21[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm28[0,0,0,1,4,4,4,5]
+; AVX512-FCP-NEXT: vpternlogd $184, %zmm2, %zmm5, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 320(%rax)
-; AVX512-FCP-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
+; AVX512-FCP-NEXT: addq $424, %rsp # imm = 0x1A8
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i8_stride6_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $328, %rsp # imm = 0x148
-; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm10
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm3
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm11
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm8
-; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm9
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm6
-; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm12
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm14 = [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-NEXT: vpbroadcastq {{.*#+}} xmm4 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm2, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm1
+; AVX512DQ-NEXT: subq $456, %rsp # imm = 0x1C8
+; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm3
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm10
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm12
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm8
+; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-NEXT: vpshufb %xmm9, %xmm3, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm9, %xmm2, %xmm1
; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm16
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm27
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm31
; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
-; AVX512DQ-NEXT: vmovdqa %xmm7, %xmm3
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm7, %xmm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm7, %xmm7
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm26
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm2, %zmm7, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm5, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm10, %xmm4
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3],xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm10, %xmm29
-; AVX512DQ-NEXT: vmovdqa64 %xmm5, %xmm30
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,3,2,1,4,5,6,7]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,6,5]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm4, %xmm2
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm4, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2
; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm11, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm11, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm11, %xmm31
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm8, %xmm4
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm8, %xmm5
+; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm27
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm8
+; AVX512DQ-NEXT: vpshufb %xmm9, %xmm6, %xmm4
+; AVX512DQ-NEXT: vpshufb %xmm9, %xmm7, %xmm5
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-NEXT: vmovdqa64 %xmm7, %xmm18
+; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm20
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,3,2,1,4,5,6,7]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,6,5]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm4
+; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm10, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm10, %xmm22
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm12, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm12, %xmm1
+; AVX512DQ-NEXT: vmovdqa64 %xmm12, %xmm21
; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm6, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm12, %ymm1
+; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm15 = [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-NEXT: vpshufb %ymm15, %ymm4, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm11, %ymm1
; AVX512DQ-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-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm12[8],ymm6[8],ymm12[9],ymm6[9],ymm12[10],ymm6[10],ymm12[11],ymm6[11],ymm12[12],ymm6[12],ymm12[13],ymm6[13],ymm12[14],ymm6[14],ymm12[15],ymm6[15],ymm12[24],ymm6[24],ymm12[25],ymm6[25],ymm12[26],ymm6[26],ymm12[27],ymm6[27],ymm12[28],ymm6[28],ymm12[29],ymm6[29],ymm12[30],ymm6[30],ymm12[31],ymm6[31]
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
-; AVX512DQ-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm1, %ymm1
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm4[8],ymm11[9],ymm4[9],ymm11[10],ymm4[10],ymm11[11],ymm4[11],ymm11[12],ymm4[12],ymm11[13],ymm4[13],ymm11[14],ymm4[14],ymm11[15],ymm4[15],ymm11[24],ymm4[24],ymm11[25],ymm4[25],ymm11[26],ymm4[26],ymm11[27],ymm4[27],ymm11[28],ymm4[28],ymm11[29],ymm4[29],ymm11[30],ymm4[30],ymm11[31],ymm4[31]
+; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
+; AVX512DQ-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm1, %ymm1
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm8, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm9, %ymm1
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm1 = 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-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15],ymm9[24],ymm8[24],ymm9[25],ymm8[25],ymm9[26],ymm8[26],ymm9[27],ymm8[27],ymm9[28],ymm8[28],ymm9[29],ymm8[29],ymm9[30],ymm8[30],ymm9[31],ymm8[31]
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
-; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm12 = [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-NEXT: vpshufb %ymm12, %ymm13, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm8, %ymm1
+; AVX512DQ-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-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm8[8],ymm13[8],ymm8[9],ymm13[9],ymm8[10],ymm13[10],ymm8[11],ymm13[11],ymm8[12],ymm13[12],ymm8[13],ymm13[13],ymm8[14],ymm13[14],ymm8[15],ymm13[15],ymm8[24],ymm13[24],ymm8[25],ymm13[25],ymm8[26],ymm13[26],ymm8[27],ymm13[27],ymm8[28],ymm13[28],ymm8[29],ymm13[29],ymm8[30],ymm13[30],ymm8[31],ymm13[31]
+; AVX512DQ-NEXT: vmovdqa64 %ymm13, %ymm17
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
+; AVX512DQ-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm1, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm13
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
+; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm13, %ymm0
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
+; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm13, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm14
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm0
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm4, %ymm10
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
-; AVX512DQ-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm11
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm3
-; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm15
-; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm15, %ymm11
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm10, %ymm7
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm11[0],ymm7[1],ymm11[1],ymm7[2],ymm11[2],ymm7[3],ymm11[3],ymm7[4],ymm11[4],ymm7[5],ymm11[5],ymm7[6],ymm11[6],ymm7[7],ymm11[7],ymm7[16],ymm11[16],ymm7[17],ymm11[17],ymm7[18],ymm11[18],ymm7[19],ymm11[19],ymm7[20],ymm11[20],ymm7[21],ymm11[21],ymm7[22],ymm11[22],ymm7[23],ymm11[23]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm10[8],ymm15[8],ymm10[9],ymm15[9],ymm10[10],ymm15[10],ymm10[11],ymm15[11],ymm10[12],ymm15[12],ymm10[13],ymm15[13],ymm10[14],ymm15[14],ymm10[15],ymm15[15],ymm10[24],ymm15[24],ymm10[25],ymm15[25],ymm10[26],ymm15[26],ymm10[27],ymm15[27],ymm10[28],ymm15[28],ymm10[29],ymm15[29],ymm10[30],ymm15[30],ymm10[31],ymm15[31]
-; AVX512DQ-NEXT: vpshufb %ymm13, %ymm11, %ymm11
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm25
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm11
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm7, %ymm13
-; AVX512DQ-NEXT: vpshufb %ymm14, %ymm11, %ymm5
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm13[0],ymm5[1],ymm13[1],ymm5[2],ymm13[2],ymm5[3],ymm13[3],ymm5[4],ymm13[4],ymm5[5],ymm13[5],ymm5[6],ymm13[6],ymm5[7],ymm13[7],ymm5[16],ymm13[16],ymm5[17],ymm13[17],ymm5[18],ymm13[18],ymm5[19],ymm13[19],ymm5[20],ymm13[20],ymm5[21],ymm13[21],ymm5[22],ymm13[22],ymm5[23],ymm13[23]
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm13 = ymm11[8],ymm7[8],ymm11[9],ymm7[9],ymm11[10],ymm7[10],ymm11[11],ymm7[11],ymm11[12],ymm7[12],ymm11[13],ymm7[13],ymm11[14],ymm7[14],ymm11[15],ymm7[15],ymm11[24],ymm7[24],ymm11[25],ymm7[25],ymm11[26],ymm7[26],ymm11[27],ymm7[27],ymm11[28],ymm7[28],ymm11[29],ymm7[29],ymm11[30],ymm7[30],ymm11[31],ymm7[31]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm24
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm23
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm7[0],ymm11[1],ymm7[1],ymm11[2],ymm7[2],ymm11[3],ymm7[3],ymm11[4],ymm7[4],ymm11[5],ymm7[5],ymm11[6],ymm7[6],ymm11[7],ymm7[7],ymm11[16],ymm7[16],ymm11[17],ymm7[17],ymm11[18],ymm7[18],ymm11[19],ymm7[19],ymm11[20],ymm7[20],ymm11[21],ymm7[21],ymm11[22],ymm7[22],ymm11[23],ymm7[23]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm1
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[2],ymm15[2],ymm10[3],ymm15[3],ymm10[4],ymm15[4],ymm10[5],ymm15[5],ymm10[6],ymm15[6],ymm10[7],ymm15[7],ymm10[16],ymm15[16],ymm10[17],ymm15[17],ymm10[18],ymm15[18],ymm10[19],ymm15[19],ymm10[20],ymm15[20],ymm10[21],ymm15[21],ymm10[22],ymm15[22],ymm10[23],ymm15[23]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm10[8],xmm2[8],xmm10[9],xmm2[9],xmm10[10],xmm2[10],xmm10[11],xmm2[11],xmm10[12],xmm2[12],xmm10[13],xmm2[13],xmm10[14],xmm2[14],xmm10[15],xmm2[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm15 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm5, %xmm0
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm28
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm1, %ymm10
+; AVX512DQ-NEXT: vpshufb %ymm15, %ymm0, %ymm15
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
+; AVX512DQ-NEXT: vpshufb %ymm6, %ymm15, %ymm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm26
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm10
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm15
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm10, %ymm6
+; AVX512DQ-NEXT: vpshufb %ymm12, %ymm15, %ymm9
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[4],ymm6[4],ymm9[5],ymm6[5],ymm9[6],ymm6[6],ymm9[7],ymm6[7],ymm9[16],ymm6[16],ymm9[17],ymm6[17],ymm9[18],ymm6[18],ymm9[19],ymm6[19],ymm9[20],ymm6[20],ymm9[21],ymm6[21],ymm9[22],ymm6[22],ymm9[23],ymm6[23]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm15[8],ymm10[8],ymm15[9],ymm10[9],ymm15[10],ymm10[10],ymm15[11],ymm10[11],ymm15[12],ymm10[12],ymm15[13],ymm10[13],ymm15[14],ymm10[14],ymm15[15],ymm10[15],ymm15[24],ymm10[24],ymm15[25],ymm10[25],ymm15[26],ymm10[26],ymm15[27],ymm10[27],ymm15[28],ymm10[28],ymm15[29],ymm10[29],ymm15[30],ymm10[30],ymm15[31],ymm10[31]
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm9, %ymm9
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm24
+; AVX512DQ-NEXT: vmovdqa (%r8), %ymm6
+; AVX512DQ-NEXT: vpshufb %ymm5, %ymm6, %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm6, %ymm4
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm23
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm4, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm4, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm25
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23]
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm3
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm15
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm15[8],xmm2[9],xmm15[9],xmm2[10],xmm15[10],xmm2[11],xmm15[11],xmm2[12],xmm15[12],xmm2[13],xmm15[13],xmm2[14],xmm15[14],xmm2[15],xmm15[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm1, %xmm10
; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm1
; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512DQ-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm29
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23]
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm30
-; AVX512DQ-NEXT: vmovdqa64 %xmm16, %xmm5
-; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm8
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm5[8],xmm8[8],xmm5[9],xmm8[9],xmm5[10],xmm8[10],xmm5[11],xmm8[11],xmm5[12],xmm8[12],xmm5[13],xmm8[13],xmm5[14],xmm8[14],xmm5[15],xmm8[15]
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm9 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512DQ-NEXT: vpshufb %xmm9, %xmm7, %xmm11
-; AVX512DQ-NEXT: vpshufb %xmm9, %xmm8, %xmm5
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm21
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm12[0],ymm6[0],ymm12[1],ymm6[1],ymm12[2],ymm6[2],ymm12[3],ymm6[3],ymm12[4],ymm6[4],ymm12[5],ymm6[5],ymm12[6],ymm6[6],ymm12[7],ymm6[7],ymm12[16],ymm6[16],ymm12[17],ymm6[17],ymm12[18],ymm6[18],ymm12[19],ymm6[19],ymm12[20],ymm6[20],ymm12[21],ymm6[21],ymm12[22],ymm6[22],ymm12[23],ymm6[23]
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm5
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm5, %xmm13
-; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm5
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm5, %xmm5
-; AVX512DQ-NEXT: vmovdqa64 %ymm5, %ymm31
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [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-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm8, %ymm3, %ymm9
-; AVX512DQ-NEXT: vpshufb %ymm8, %ymm4, %ymm3
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm9, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm20
+; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[1],ymm3[1],ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[4],ymm3[4],ymm8[5],ymm3[5],ymm8[6],ymm3[6],ymm8[7],ymm3[7],ymm8[16],ymm3[16],ymm8[17],ymm3[17],ymm8[18],ymm3[18],ymm8[19],ymm3[19],ymm8[20],ymm3[20],ymm8[21],ymm3[21],ymm8[22],ymm3[22],ymm8[23],ymm3[23]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm17
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm7
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm3[8],xmm7[9],xmm3[9],xmm7[10],xmm3[10],xmm7[11],xmm3[11],xmm7[12],xmm3[12],xmm7[13],xmm3[13],xmm7[14],xmm3[14],xmm7[15],xmm3[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512DQ-NEXT: vpshufb %xmm8, %xmm5, %xmm9
+; AVX512DQ-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm11[0],ymm3[0],ymm11[1],ymm3[1],ymm11[2],ymm3[2],ymm11[3],ymm3[3],ymm11[4],ymm3[4],ymm11[5],ymm3[5],ymm11[6],ymm3[6],ymm11[7],ymm3[7],ymm11[16],ymm3[16],ymm11[17],ymm3[17],ymm11[18],ymm3[18],ymm11[19],ymm3[19],ymm11[20],ymm3[20],ymm11[21],ymm3[21],ymm11[22],ymm3[22],ymm11[23],ymm3[23]
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
+; AVX512DQ-NEXT: vmovdqa64 %xmm22, %xmm5
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm5, %xmm7
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm5
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm5, %xmm3
; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm1, %xmm3
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm0, %xmm4
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [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-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm6, %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm13, %ymm6
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm13
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm11
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm4, %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm3
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm1, %xmm14
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm8
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; AVX512DQ-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm27
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm2, %xmm0
-; AVX512DQ-NEXT: vpshufb %xmm14, %xmm10, %xmm1
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm1
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3],xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm15, %xmm1
+; AVX512DQ-NEXT: vpshufb %xmm12, %xmm2, %xmm8
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7]
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm1, %xmm6
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm15
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm3, %xmm10
-; AVX512DQ-NEXT: vmovdqa64 %ymm10, %ymm17
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm10
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm14
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm1, %xmm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm1, %ymm5
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm3, %ymm2
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
-; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm7
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
-; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm12
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm1, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm1, %ymm4
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm18[0,0,0,1]
-; AVX512DQ-NEXT: vprold $16, %ymm28, %ymm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm8
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm10[0,0,0,1]
+; AVX512DQ-NEXT: vprold $16, %ymm16, %ymm8
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm6[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm13[0,0,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm20[0,0,0,1]
+; AVX512DQ-NEXT: vprold $16, %ymm17, %ymm14
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm31[0,0,0,1]
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm6, %zmm1
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512DQ-NEXT: vpternlogq $184, %ymm0, %ymm6, %ymm5
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512DQ-NEXT: vpternlogq $184, %ymm1, %ymm0, %ymm28
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm1
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm28[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm13, %zmm1
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm5, %zmm8, %zmm1
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm29[0,0,0,1]
-; AVX512DQ-NEXT: vprold $16, %ymm30, %ymm9
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm21[0,0,0,1]
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm13
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm28 = ymm31[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm29 = ymm22[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm30 = ymm17[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm16[0,0,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm22[0,0,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm9
-; AVX512DQ-NEXT: vpternlogq $226, %zmm5, %zmm6, %zmm9
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm9, %ymm5
-; AVX512DQ-NEXT: vpternlogq $184, %ymm5, %ymm6, %ymm2
-; AVX512DQ-NEXT: vpternlogq $184, %ymm9, %ymm0, %ymm15
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm2
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm15[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm5
-; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm8, %zmm5
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm27[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm8
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm9, %zmm8
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm8, %ymm2
-; AVX512DQ-NEXT: vpternlogq $184, %ymm2, %ymm9, %ymm30
-; AVX512DQ-NEXT: vpternlogq $184, %ymm8, %ymm6, %ymm10
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm2
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm10[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm10, %zmm8
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm26[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm11 = mem[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm9, %zmm11
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm11, %ymm2
-; AVX512DQ-NEXT: vpternlogq $184, %ymm2, %ymm9, %ymm14
-; AVX512DQ-NEXT: vpternlogq $184, %ymm11, %ymm6, %ymm31
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm2
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm31[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm11 = mem[0,0,0,1,4,4,4,5]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm10, %zmm11
-; AVX512DQ-NEXT: vpermq $234, (%rsp), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm10 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm10
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm25[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm13 = zmm24[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm6, %zmm13
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm10, %ymm2
-; AVX512DQ-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm7
-; AVX512DQ-NEXT: vextracti64x4 $1, %zmm13, %ymm2
-; AVX512DQ-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm3
-; AVX512DQ-NEXT: vpternlogq $184, %ymm10, %ymm9, %ymm12
-; AVX512DQ-NEXT: vpternlogq $184, %ymm13, %ymm9, %ymm4
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm12[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm3 = zmm23[2,2,2,3,6,6,6,7]
-; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm6, %zmm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm2
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm8, %zmm5, %zmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm14[2,2,2,3]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm1
+; AVX512DQ-NEXT: vpternlogq $226, %zmm7, %zmm9, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6
+; AVX512DQ-NEXT: vpternlogd $184, %zmm1, %zmm5, %zmm6
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm12, %zmm1
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm2
+; AVX512DQ-NEXT: vpternlogd $184, %zmm6, %zmm4, %zmm2
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm4, %zmm3
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm3, %zmm9, %zmm0
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm4, %zmm6
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm6, %zmm9, %zmm0
+; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm6
+; AVX512DQ-NEXT: vpermq $234, (%rsp), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm7
+; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm5, %zmm4, %zmm0
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm23[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm7, %zmm4, %zmm5
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm4 = zmm28[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm7 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm0, %zmm7, %zmm4
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm25[2,2,2,3,6,6,6,7]
+; AVX512DQ-NEXT: vpternlogd $184, %zmm5, %zmm7, %zmm0
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm3, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm11, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm8, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm4, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm6, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm3, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, 256(%rax)
; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-NEXT: addq $328, %rsp # imm = 0x148
+; AVX512DQ-NEXT: addq $456, %rsp # imm = 0x1C8
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i8_stride6_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $392, %rsp # imm = 0x188
+; AVX512DQ-FCP-NEXT: subq $424, %rsp # imm = 0x1A8
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-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-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm1
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm2
; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm2 = 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-FCP-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15],ymm3[24],ymm4[24],ymm3[25],ymm4[25],ymm3[26],ymm4[26],ymm3[27],ymm4[27],ymm3[28],ymm4[28],ymm3[29],ymm4[29],ymm3[30],ymm4[30],ymm3[31],ymm4[31]
@@ -6276,258 +6226,238 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm5
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm14 = [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-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm10 = [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-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[16],ymm2[16],ymm3[17],ymm2[17],ymm3[18],ymm2[18],ymm3[19],ymm2[19],ymm3[20],ymm2[20],ymm3[21],ymm2[21],ymm3[22],ymm2[22],ymm3[23],ymm2[23]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15],ymm6[24],ymm5[24],ymm6[25],ymm5[25],ymm6[26],ymm5[26],ymm6[27],ymm5[27],ymm6[28],ymm5[28],ymm6[29],ymm5[29],ymm6[30],ymm5[30],ymm6[31],ymm5[31]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm25
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm30
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15],ymm4[24],ymm5[24],ymm4[25],ymm5[25],ymm4[26],ymm5[26],ymm4[27],ymm5[27],ymm4[28],ymm5[28],ymm4[29],ymm5[29],ymm4[30],ymm5[30],ymm4[31],ymm5[31]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm22
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm4, %ymm4
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm7
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm5
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm5
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm7, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm31
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm6
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[4],ymm5[4],ymm0[5],ymm5[5],ymm0[6],ymm5[6],ymm0[7],ymm5[7],ymm0[16],ymm5[16],ymm0[17],ymm5[17],ymm0[18],ymm5[18],ymm0[19],ymm5[19],ymm0[20],ymm5[20],ymm0[21],ymm5[21],ymm0[22],ymm5[22],ymm0[23],ymm5[23]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm7[8],ymm8[8],ymm7[9],ymm8[9],ymm7[10],ymm8[10],ymm7[11],ymm8[11],ymm7[12],ymm8[12],ymm7[13],ymm8[13],ymm7[14],ymm8[14],ymm7[15],ymm8[15],ymm7[24],ymm8[24],ymm7[25],ymm8[25],ymm7[26],ymm8[26],ymm7[27],ymm8[27],ymm7[28],ymm8[28],ymm7[29],ymm8[29],ymm7[30],ymm8[30],ymm7[31],ymm8[31]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
+; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
+; AVX512DQ-FCP-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[1],ymm6[1],ymm0[2],ymm6[2],ymm0[3],ymm6[3],ymm0[4],ymm6[4],ymm0[5],ymm6[5],ymm0[6],ymm6[6],ymm0[7],ymm6[7],ymm0[16],ymm6[16],ymm0[17],ymm6[17],ymm0[18],ymm6[18],ymm0[19],ymm6[19],ymm0[20],ymm6[20],ymm0[21],ymm6[21],ymm0[22],ymm6[22],ymm0[23],ymm6[23]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15],ymm8[24],ymm7[24],ymm8[25],ymm7[25],ymm8[26],ymm7[26],ymm8[27],ymm7[27],ymm8[28],ymm7[28],ymm8[29],ymm7[29],ymm8[30],ymm7[30],ymm8[31],ymm7[31]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm8, %ymm18
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm17
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm1
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm11, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm12, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm15, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm11, %ymm1
; AVX512DQ-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-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm12[8],ymm11[8],ymm12[9],ymm11[9],ymm12[10],ymm11[10],ymm12[11],ymm11[11],ymm12[12],ymm11[12],ymm12[13],ymm11[13],ymm12[14],ymm11[14],ymm12[15],ymm11[15],ymm12[24],ymm11[24],ymm12[25],ymm11[25],ymm12[26],ymm11[26],ymm12[27],ymm11[27],ymm12[28],ymm11[28],ymm12[29],ymm11[29],ymm12[30],ymm11[30],ymm12[31],ymm11[31]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm15[8],ymm11[9],ymm15[9],ymm11[10],ymm15[10],ymm11[11],ymm15[11],ymm11[12],ymm15[12],ymm11[13],ymm15[13],ymm11[14],ymm15[14],ymm11[15],ymm15[15],ymm11[24],ymm15[24],ymm11[25],ymm15[25],ymm11[26],ymm15[26],ymm11[27],ymm15[27],ymm11[28],ymm15[28],ymm11[29],ymm15[29],ymm11[30],ymm15[30],ymm11[31],ymm15[31]
; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm3
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm23
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm12, %xmm2
; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3],xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm18
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm4, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm19
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm5, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm27
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm15
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm15, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm15, %xmm1
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm28
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm13
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm2, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm10
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3],xmm12[4],xmm0[4],xmm12[5],xmm0[5],xmm12[6],xmm0[6],xmm12[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm16
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm26
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm13, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm13, %xmm3
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm29
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm14
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm14, %xmm4
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm4, %zmm21
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm4
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm4, %xmm5
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm5, %xmm5
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm5, %zmm25
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm2
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm2, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm3
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm5, %zmm3, %zmm28
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm3, %xmm9
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm30
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm7, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3],xmm10[4],xmm0[4],xmm10[5],xmm0[5],xmm10[6],xmm0[6],xmm10[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX512DQ-FCP-NEXT: vprold $16, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm10, %zmm23
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm4, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm10, %xmm14
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm14, %xmm14
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm29
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm12[0],ymm11[0],ymm12[1],ymm11[1],ymm12[2],ymm11[2],ymm12[3],ymm11[3],ymm12[4],ymm11[4],ymm12[5],ymm11[5],ymm12[6],ymm11[6],ymm12[7],ymm11[7],ymm12[16],ymm11[16],ymm12[17],ymm11[17],ymm12[18],ymm11[18],ymm12[19],ymm11[19],ymm12[20],ymm11[20],ymm12[21],ymm11[21],ymm12[22],ymm11[22],ymm12[23],ymm11[23]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm31
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm15[0],ymm11[1],ymm15[1],ymm11[2],ymm15[2],ymm11[3],ymm15[3],ymm11[4],ymm15[4],ymm11[5],ymm15[5],ymm11[6],ymm15[6],ymm11[7],ymm15[7],ymm11[16],ymm15[16],ymm11[17],ymm15[17],ymm11[18],ymm15[18],ymm11[19],ymm15[19],ymm11[20],ymm15[20],ymm11[21],ymm15[21],ymm11[22],ymm15[22],ymm11[23],ymm15[23]
; AVX512DQ-FCP-NEXT: vprold $16, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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: vpunpcklbw {{.*#+}} ymm4 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[4],ymm8[4],ymm7[5],ymm8[5],ymm7[6],ymm8[6],ymm7[7],ymm8[7],ymm7[16],ymm8[16],ymm7[17],ymm8[17],ymm7[18],ymm8[18],ymm7[19],ymm8[19],ymm7[20],ymm8[20],ymm7[21],ymm8[21],ymm7[22],ymm8[22],ymm7[23],ymm8[23]
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
-; AVX512DQ-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm11
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,0,0,1,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm10, %xmm3
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm16, %zmm11
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [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-FCP-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm15, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm5, %xmm17
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm5
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[4],ymm5[4],ymm0[5],ymm5[5],ymm0[6],ymm5[6],ymm0[7],ymm5[7],ymm0[16],ymm5[16],ymm0[17],ymm5[17],ymm0[18],ymm5[18],ymm0[19],ymm5[19],ymm0[20],ymm5[20],ymm0[21],ymm5[21],ymm0[22],ymm5[22],ymm0[23],ymm5[23]
-; AVX512DQ-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-FCP-NEXT: vprold $16, %ymm0, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm1, %xmm2
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm4 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[16],mem[16],ymm0[17],mem[17],ymm0[18],mem[18],ymm0[19],mem[19],ymm0[20],mem[20],ymm0[21],mem[21],ymm0[22],mem[22],ymm0[23],mem[23]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm14
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm8
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm15
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm8
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm15, %xmm4
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm8
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm9
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm2[0],ymm9[0],ymm2[1],ymm9[1],ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[4],ymm9[4],ymm2[5],ymm9[5],ymm2[6],ymm9[6],ymm2[7],ymm9[7],ymm2[16],ymm9[16],ymm2[17],ymm9[17],ymm2[18],ymm9[18],ymm2[19],ymm9[19],ymm2[20],ymm9[20],ymm2[21],ymm9[21],ymm2[22],ymm9[22],ymm2[23],ymm9[23]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15]
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
+; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm4, %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm17
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,0,0,1,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm4, %zmm15
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm7, %zmm9, %zmm15
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm14, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm15, %xmm5
-; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm14, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm15, %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm15, %ymm15
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm30 = ymm0[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm31 = ymm1[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm6[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm14[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm11, %ymm14, %ymm2
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm11, %ymm11
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm8, %ymm14, %ymm13
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm8
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm8, %ymm16, %ymm9
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm7
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm13[0,1,2,3],zmm7[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm7, %zmm8, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm15, %zmm10, %zmm1
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm8, %xmm0
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm1, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm8
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0],ymm8[0],ymm1[1],ymm8[1],ymm1[2],ymm8[2],ymm1[3],ymm8[3],ymm1[4],ymm8[4],ymm1[5],ymm8[5],ymm1[6],ymm8[6],ymm1[7],ymm8[7],ymm1[16],ymm8[16],ymm1[17],ymm8[17],ymm1[18],ymm8[18],ymm1[19],ymm8[19],ymm1[20],ymm8[20],ymm1[21],ymm8[21],ymm1[22],ymm8[22],ymm1[23],ymm8[23]
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vmovdqu (%rsp), %ymm5 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm5 = ymm5[0],mem[0],ymm5[1],mem[1],ymm5[2],mem[2],ymm5[3],mem[3],ymm5[4],mem[4],ymm5[5],mem[5],ymm5[6],mem[6],ymm5[7],mem[7],ymm5[16],mem[16],ymm5[17],mem[17],ymm5[18],mem[18],ymm5[19],mem[19],ymm5[20],mem[20],ymm5[21],mem[21],ymm5[22],mem[22],ymm5[23],mem[23]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm6
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm12[8],xmm6[8],xmm12[9],xmm6[9],xmm12[10],xmm6[10],xmm12[11],xmm6[11],xmm12[12],xmm6[12],xmm12[13],xmm6[13],xmm12[14],xmm6[14],xmm12[15],xmm6[15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm6
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm9, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm13, %xmm3
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm6, %zmm10, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm5
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm2, %zmm5
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax)
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm11, %ymm16, %ymm12
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm6
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm8, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
-; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm3 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm3
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm3, %ymm2
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm2, %ymm14, %ymm22
-; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm6
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm6, %ymm2
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm2, %ymm14, %ymm30
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm1 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm1
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm24
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm24[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm7 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm8, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm6, %ymm2, %ymm31
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm31[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $234, (%rsp), %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm3, %zmm8, %zmm6
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm23[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm8
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm3
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm4
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm29[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm27[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm9
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm9, %ymm3
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm3, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm8, %ymm16, %ymm5
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %ymm9, %ymm16, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm2
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm5[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm1 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm1
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm4 = mem[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm4
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm23[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm4, %zmm2, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm24[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm3, %zmm4
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5]
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm26[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm4, %zmm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm28[0,0,0,1,4,4,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm4, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, (%rax)
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm31[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm25[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm29[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm3, %zmm9, %zmm0
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm27[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm5, %zmm9, %zmm2
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm21[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm0, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm28[0,0,0,1,4,4,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogd $184, %zmm2, %zmm5, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 320(%rax)
-; AVX512DQ-FCP-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
+; AVX512DQ-FCP-NEXT: addq $424, %rsp # imm = 0x1A8
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
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 8091afbbfd70c..416fbe9aa340c 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
@@ -4254,7 +4254,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512-NEXT: vmovdqa64 %ymm19, %ymm14
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u,u,u,u,u,14],zero,ymm14[u,u,u,u,u,15],zero,ymm14[u,u,u,u,u,16],zero,ymm14[u,u,u,u,u,17],zero,ymm14[u,u,u]
+; AVX512-NEXT: vpshufb {{.*#+}} 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]
@@ -4357,7 +4357,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm9
; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u,u,u,u,u,14],zero,ymm1[u,u,u,u,u,15],zero,ymm1[u,u,u,u,u,16],zero,ymm1[u,u,u,u,u,17],zero,ymm1[u,u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u],zero,zero,zero,zero,ymm1[14],zero,ymm1[u],zero,zero,zero,zero,ymm1[15],zero,ymm1[u],zero,zero,zero,zero,ymm1[16],zero,ymm1[u],zero,zero,zero,zero,ymm1[17],zero,ymm1[u],zero,zero
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
; AVX512-FCP-NEXT: vpor %ymm0, %ymm8, %ymm0
; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
@@ -4533,7 +4533,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm14
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm14[u,u,u,u,u,14],zero,ymm14[u,u,u,u,u,15],zero,ymm14[u,u,u,u,u,16],zero,ymm14[u,u,u,u,u,17],zero,ymm14[u,u,u]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} 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]
@@ -4636,7 +4636,7 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm9
; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u,u,u,u,u,14],zero,ymm1[u,u,u,u,u,15],zero,ymm1[u,u,u,u,u,16],zero,ymm1[u,u,u,u,u,17],zero,ymm1[u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm1[u],zero,zero,zero,zero,ymm1[14],zero,ymm1[u],zero,zero,zero,zero,ymm1[15],zero,ymm1[u],zero,zero,zero,zero,ymm1[16],zero,ymm1[u],zero,zero,zero,zero,ymm1[17],zero,ymm1[u],zero,zero
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm2[13,u,u,u,u,u],zero,ymm2[14,u,u,u,u,u],zero,ymm2[15,u,u,u,u,u],zero,ymm2[16,u,u,u,u,u],zero,ymm2[17,u,u,u]
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm8, %ymm0
; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
@@ -8544,39 +8544,35 @@ 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 $1720, %rsp # imm = 0x6B8
+; AVX512-NEXT: subq $1384, %rsp # imm = 0x568
; AVX512-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512-NEXT: vpshufb %ymm2, %ymm7, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero,zero,zero,ymm7[18]
; AVX512-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm25
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm20
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%rcx), %ymm14
+; AVX512-NEXT: vmovdqa (%rcx), %ymm15
; AVX512-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512-NEXT: vpshufb %ymm2, %ymm15, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm23
+; AVX512-NEXT: vmovdqa (%rdx), %ymm2
; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512-NEXT: vpshufb %ymm3, %ymm8, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm23
-; AVX512-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa (%r8), %ymm15
+; 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, %ymm15, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm24
-; AVX512-NEXT: vmovdqa (%r9), %ymm2
+; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-NEXT: vmovdqa (%r9), %ymm8
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
; AVX512-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm3, %ymm27
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-NEXT: vpshufb %ymm3, %ymm8, %ymm1
+; AVX512-NEXT: vmovdqa64 %ymm3, %ymm17
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-NEXT: vmovdqa 32(%rsi), %ymm10
@@ -8589,7 +8585,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,14,128,12,128,14,15,14,15,128,13,128,15,12,13,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm19
; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm10[23,u,u,u],zero,ymm10[26],zero,ymm10[24,u,u,u],zero,ymm10[27],zero
; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -8604,7 +8600,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0]
; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm2
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm21
; AVX512-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -8612,7 +8608,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
; AVX512-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm11
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512-NEXT: vmovdqa64 %ymm0, %ymm25
; AVX512-NEXT: vmovdqa 32(%r9), %ymm2
; AVX512-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
@@ -8631,23 +8627,23 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vpshufb %ymm6, %ymm7, %ymm6
-; AVX512-NEXT: vmovdqa64 %ymm7, %ymm17
-; AVX512-NEXT: vmovdqa64 %ymm18, %ymm13
-; AVX512-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm13, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm7, %ymm22
+; AVX512-NEXT: vmovdqa64 %ymm20, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm7, %ymm7
; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %ymm1, %ymm8, %ymm6
-; AVX512-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm14, %ymm7
-; AVX512-NEXT: vmovdqa64 %ymm14, %ymm22
+; AVX512-NEXT: vmovdqa64 %ymm18, %ymm12
+; AVX512-NEXT: vpshufb %ymm1, %ymm12, %ymm6
+; AVX512-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm15, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm15, %ymm19
; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm15, %ymm6
-; AVX512-NEXT: vmovdqa64 %ymm15, %ymm16
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm15
-; AVX512-NEXT: vpshufb %ymm11, %ymm15, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512-NEXT: vpshufb %ymm11, %ymm8, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm8, %ymm16
; AVX512-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0
@@ -8656,379 +8652,348 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
; AVX512-NEXT: vpshufb %xmm11, %xmm1, %xmm6
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm7
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm7
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm21
; AVX512-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512-NEXT: vmovdqa 32(%rsi), %xmm15
; AVX512-NEXT: vmovdqa {{.*#+}} xmm6 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512-NEXT: vpshufb %xmm6, %xmm14, %xmm7
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm8
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512-NEXT: vpshufb %xmm6, %xmm15, %xmm7
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm13 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512-NEXT: vpshufb %xmm13, %xmm0, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm30
; AVX512-NEXT: vpor %xmm7, %xmm8, %xmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
; AVX512-NEXT: vmovdqa 32(%rax), %xmm0
; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT: vpshufhw {{.*#+}} xmm7 = xmm0[0,1,2,3,4,5,5,6]
; AVX512-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpermi2d %zmm7, %zmm8, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vpermi2d %zmm7, %zmm8, %zmm28
; AVX512-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa 32(%r8), %xmm1
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512-NEXT: vpshufb %xmm8, %xmm0, %xmm7
-; AVX512-NEXT: vmovdqa64 %xmm8, %xmm20
+; AVX512-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm7
+; AVX512-NEXT: vmovdqa64 %xmm1, %xmm24
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm18
; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm30
-; AVX512-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512-NEXT: vpor %xmm7, %xmm8, %xmm0
-; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm28, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm7
+; AVX512-NEXT: vpshufb %xmm0, %xmm14, %xmm8
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm29
+; AVX512-NEXT: vporq %xmm7, %xmm8, %xmm31
; AVX512-NEXT: vmovdqa64 %ymm23, %ymm0
+; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm7
+; AVX512-NEXT: vmovdqa64 %ymm26, %ymm0
; AVX512-NEXT: vpshufb %ymm0, %ymm5, %ymm8
; AVX512-NEXT: vpor %ymm7, %ymm8, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
; AVX512-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512-NEXT: vpshufb %ymm1, %ymm4, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
-; AVX512-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm4, %ymm5, %ymm1
-; AVX512-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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: # ymm7 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm7, %ymm5, %ymm1
; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm26, %ymm0
-; AVX512-NEXT: vpshufb %ymm0, %ymm10, %ymm0
-; AVX512-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm1
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm10[14],zero,zero,zero,zero,zero,zero,ymm10[15],zero,zero,zero,zero,zero,zero,ymm10[16],zero,zero,zero,zero,zero,zero,ymm10[17],zero,zero,zero,zero,zero,zero,ymm10[18]
+; AVX512-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-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX512-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm7, %ymm10, %ymm0
-; AVX512-NEXT: vmovdqa {{.*#+}} ymm1 = [2,3,4,5,128,3,128,5,4,5,6,128,4,128,6,7,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512-NEXT: vpshufb %ymm1, %ymm9, %ymm4
-; AVX512-NEXT: vpor %ymm0, %ymm4, %ymm0
-; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm10, %ymm2, %ymm4
-; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512-NEXT: 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-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm5
+; AVX512-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512-NEXT: vmovdqa {{.*#+}} ymm8 = [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 %ymm8, %ymm9, %ymm4
+; AVX512-NEXT: vpor %ymm1, %ymm4, %ymm1
+; AVX512-NEXT: vmovdqu %ymm1, {{[-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, %ymm4
+; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
+; AVX512-NEXT: vpshufb %ymm9, %ymm3, %ymm5
; AVX512-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm24, %ymm4
-; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm3
; AVX512-NEXT: vmovdqa64 %ymm27, %ymm4
+; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX512-NEXT: vmovdqa64 %ymm17, %ymm4
; AVX512-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512-NEXT: vpor %ymm3, %ymm2, %ymm2
; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT: vmovdqa (%rsi), %xmm3
; AVX512-NEXT: vpshufb %xmm6, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512-NEXT: vmovdqa64 %xmm3, %xmm25
; AVX512-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512-NEXT: vmovdqa64 %xmm18, %xmm3
-; AVX512-NEXT: vpshufb %xmm3, %xmm4, %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512-NEXT: vpshufb %xmm13, %xmm4, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm4, %xmm27
; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT: vmovdqa (%rcx), %xmm3
; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa %xmm3, %xmm6
-; AVX512-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512-NEXT: vpshufb %xmm12, %xmm4, %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512-NEXT: vmovdqa64 %xmm3, %xmm17
+; AVX512-NEXT: vmovdqa (%rdx), %xmm13
+; AVX512-NEXT: vmovdqa64 %xmm21, %xmm3
+; AVX512-NEXT: vpshufb %xmm3, %xmm13, %xmm3
; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-NEXT: vmovdqa (%r9), %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm3, %xmm2
-; AVX512-NEXT: vmovdqa %xmm3, %xmm8
-; AVX512-NEXT: vmovdqa (%r8), %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm30, %xmm4
-; AVX512-NEXT: vpshufb %xmm4, %xmm3, %xmm4
-; AVX512-NEXT: vmovdqa %xmm3, %xmm9
-; AVX512-NEXT: vpor %xmm2, %xmm4, %xmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero,ymm13[27],zero,ymm13[25]
-; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512-NEXT: vpshufb %ymm7, %ymm3, %ymm4
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23,u,u,u],zero,ymm3[26],zero,ymm3[24,u,u,u],zero,ymm3[27],zero
-; AVX512-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX512-NEXT: vmovdqa64 %ymm21, %ymm3
-; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
-; AVX512-NEXT: vmovdqa64 %ymm19, %ymm3
-; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
-; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm15[25],zero,ymm15[23],zero,zero,zero,zero,ymm15[26],zero,ymm15[24],zero,zero
-; AVX512-NEXT: vmovdqa64 %ymm16, %ymm3
-; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
-; AVX512-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vpshufb %ymm10, %ymm15, %ymm1
+; AVX512-NEXT: vmovdqa (%r9), %xmm11
+; AVX512-NEXT: vmovdqa64 %xmm24, %xmm2
+; AVX512-NEXT: vpshufb %xmm2, %xmm11, %xmm2
+; AVX512-NEXT: vmovdqa (%r8), %xmm10
+; AVX512-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512-NEXT: vpshufb %xmm3, %xmm10, %xmm3
+; AVX512-NEXT: vpor %xmm2, %xmm3, %xmm2
+; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-NEXT: vmovdqa64 %ymm20, %ymm4
+; AVX512-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero,ymm4[27],zero,ymm4[25]
+; AVX512-NEXT: vmovdqa64 %ymm22, %ymm3
; AVX512-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-NEXT: vmovdqa (%rax), %ymm4
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm15
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512-NEXT: vmovdqa64 %xmm9, %xmm21
-; AVX512-NEXT: vmovdqa64 %xmm8, %xmm22
-; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm2, %xmm24
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm1[0,1,0,1],zmm0[4,5,6,7]
-; AVX512-NEXT: vmovdqa (%rax), %xmm11
-; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,5,5,6]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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 %ymm8, %ymm4, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm29
+; AVX512-NEXT: vmovdqa64 %ymm19, %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 %ymm23, %ymm3
+; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb %ymm7, %ymm12, %ymm2
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm23
+; 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,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; 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 %ymm9, %ymm1, %ymm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-NEXT: vmovdqa (%rax), %ymm8
+; AVX512-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm2 = ymm8[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm16
+; AVX512-NEXT: vmovdqa (%rax), %xmm5
+; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm5[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 {{.*#+}} ymm9 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
-; AVX512-NEXT: vpshufb %ymm9, %ymm4, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3],xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512-NEXT: vpshufb %ymm9, %ymm8, %ymm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm19
+; AVX512-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm15[0],xmm4[1],xmm15[1],xmm4[2],xmm15[2],xmm4[3],xmm15[3],xmm4[4],xmm15[4],xmm4[5],xmm15[5],xmm4[6],xmm15[6],xmm4[7],xmm15[7]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512-NEXT: vpshufb %xmm0, %xmm2, %xmm2
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm24
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[2,3,2,3],zmm2[0,1,0,1]
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[2,3,2,3],zmm3[0,1,0,1]
; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
; AVX512-NEXT: vpshufb %xmm0, %xmm3, %xmm3
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm26
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm21
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[2,3,2,3],zmm3[0,1,0,1]
-; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512-NEXT: vmovdqa64 %xmm29, %xmm3
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm14[0],xmm2[0],xmm14[1],xmm2[1],xmm14[2],xmm2[2],xmm14[3],xmm2[3],xmm14[4],xmm2[4],xmm14[5],xmm2[5],xmm14[6],xmm2[6],xmm14[7],xmm2[7]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512-NEXT: vmovdqa64 %xmm0, %xmm25
+; AVX512-NEXT: vpshufb %xmm0, %xmm12, %xmm12
+; AVX512-NEXT: vmovdqa64 %xmm0, %xmm20
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[2,3,2,3],zmm13[0,1,0,1]
-; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512-NEXT: vmovdqa %xmm6, %xmm8
-; AVX512-NEXT: vmovdqa64 %xmm18, %xmm7
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
+; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[2,3,2,3],zmm12[0,1,0,1]
+; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[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, %ymm18
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm7[8],xmm1[8],xmm7[9],xmm1[9],xmm7[10],xmm1[10],xmm7[11],xmm1[11],xmm7[12],xmm1[12],xmm7[13],xmm1[13],xmm7[14],xmm1[14],xmm7[15],xmm1[15]
+; AVX512-NEXT: vmovdqa64 %xmm17, %xmm3
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm3[8],xmm13[8],xmm3[9],xmm13[9],xmm3[10],xmm13[10],xmm3[11],xmm13[11],xmm3[12],xmm13[12],xmm3[13],xmm13[13],xmm3[14],xmm13[14],xmm3[15],xmm13[15]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-NEXT: vpshufb %xmm1, %xmm4, %xmm2
-; AVX512-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512-NEXT: vpshufb %xmm1, %xmm8, %xmm12
; AVX512-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
-; AVX512-NEXT: vmovdqa64 %xmm31, %xmm6
-; AVX512-NEXT: vmovdqa64 %xmm28, %xmm4
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15]
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm8
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm4[8],xmm15[9],xmm4[9],xmm15[10],xmm4[10],xmm15[11],xmm4[11],xmm15[12],xmm4[12],xmm15[13],xmm4[13],xmm15[14],xmm4[14],xmm15[15],xmm4[15]
+; AVX512-NEXT: vmovdqa64 %xmm25, %xmm4
+; AVX512-NEXT: vmovdqa64 %xmm27, %xmm7
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm4[8],xmm7[8],xmm4[9],xmm7[9],xmm4[10],xmm7[10],xmm4[11],xmm7[11],xmm4[12],xmm7[12],xmm4[13],xmm7[13],xmm4[14],xmm7[14],xmm4[15],xmm7[15]
; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-NEXT: vpshufb %xmm0, %xmm14, %xmm13
+; AVX512-NEXT: vpshufb %xmm0, %xmm15, %xmm15
; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm0
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
-; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
-; AVX512-NEXT: vmovdqa64 %xmm24, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512-NEXT: vpshufb %ymm9, %ymm10, %ymm1
-; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm17
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm2[8],xmm14[9],xmm2[9],xmm14[10],xmm2[10],xmm14[11],xmm2[11],xmm14[12],xmm2[12],xmm14[13],xmm2[13],xmm14[14],xmm2[14],xmm14[15],xmm2[15]
+; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm10[8],xmm11[8],xmm10[9],xmm11[9],xmm10[10],xmm11[10],xmm10[11],xmm11[11],xmm10[12],xmm11[12],xmm10[13],xmm11[13],xmm10[14],xmm11[14],xmm10[15],xmm11[15]
+; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-NEXT: vpshufb %xmm0, %xmm14, %xmm14
+; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm31, %zmm31
+; AVX512-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-NEXT: vpshufb %ymm9, %ymm0, %ymm1
+; AVX512-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512-NEXT: vpternlogq $226, %zmm12, %zmm10, %zmm17
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm2 = mem[2,3,2,3]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
-; AVX512-NEXT: # ymm28 = mem[2,3,2,3]
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-NEXT: vinserti64x4 $1, %ymm28, %zmm3, %zmm28
-; AVX512-NEXT: vpternlogq $226, %zmm2, %zmm10, %zmm28
-; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm9[2,2,3,2]
-; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm10
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512-NEXT: vmovdqa64 %xmm27, %xmm2
-; AVX512-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm27 # 16-byte Folded Reload
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
-; AVX512-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512-NEXT: vpshufb %xmm1, %xmm2, %xmm2
-; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
-; AVX512-NEXT: vmovdqa64 %xmm21, %xmm1
-; AVX512-NEXT: vmovdqa64 %xmm22, %xmm3
-; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512-NEXT: vmovdqa64 %xmm25, %xmm1
-; AVX512-NEXT: vpshufb %xmm1, %xmm8, %xmm8
-; AVX512-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
-; AVX512-NEXT: vpshufb {{.*#+}} xmm9 = xmm11[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm11[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,2]
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm9
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm1 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512-NEXT: vpternlogq $226, %zmm6, %zmm1, %zmm22
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm6 = mem[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512-NEXT: # ymm27 = mem[2,3,2,3]
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm27
+; AVX512-NEXT: vpternlogq $226, %zmm6, %zmm1, %zmm27
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3],xmm7[4],xmm4[4],xmm7[5],xmm4[5],xmm7[6],xmm4[6],xmm7[7],xmm4[7]
+; AVX512-NEXT: vmovdqa64 %xmm24, %xmm0
+; AVX512-NEXT: vpshufb %xmm0, %xmm1, %xmm1
+; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm13[0],xmm3[0],xmm13[1],xmm3[1],xmm13[2],xmm3[2],xmm13[3],xmm3[3],xmm13[4],xmm3[4],xmm13[5],xmm3[5],xmm13[6],xmm3[6],xmm13[7],xmm3[7]
+; AVX512-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512-NEXT: vpshufb %xmm0, %xmm6, %xmm6
+; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 16-byte Folded Reload
+; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3],xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
+; AVX512-NEXT: vmovdqa64 %xmm20, %xmm0
+; AVX512-NEXT: vpshufb %xmm0, %xmm10, %xmm10
+; AVX512-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 16-byte Folded Reload
+; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm5[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[1,1,0,0,4,5,6,7]
; AVX512-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm9, %zmm11, %zmm9
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm11, %zmm12, %zmm11
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm11
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm9, %zmm12, %zmm9
-; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512-NEXT: vpternlogq $184, %zmm11, %zmm12, %zmm9
-; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm13[0,1,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm18[0,1,0,1]
-; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,3,2,3]
-; AVX512-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; AVX512-NEXT: # xmm1 = mem[1,1,0,0,4,5,6,7]
-; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
-; AVX512-NEXT: vpermq {{.*#+}} zmm13 = zmm19[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm0
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm21 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm13, %zmm21, %zmm13
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm20 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpermq {{.*#+}} zmm18 = zmm30[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vporq %zmm20, %zmm18, %zmm18
-; AVX512-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm18
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
-; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm4
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm16
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
-; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
-; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,0]
-; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm1
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm1
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm5
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm13 # 64-byte Folded Reload
+; AVX512-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm13 = zmm13[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm13
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm21 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512-NEXT: vpternlogq $184, %zmm13, %zmm21, %zmm11
+; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm17[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpternlogq $226, %zmm8, %zmm21, %zmm7
+; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm15[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm14[0,1,0,1]
+; AVX512-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,3,2,3]
+; AVX512-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512-NEXT: # xmm15 = mem[1,1,0,0,4,5,6,7]
+; AVX512-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
+; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm20 # 64-byte Folded Reload
+; AVX512-NEXT: vporq %zmm26, %zmm23, %zmm17
+; AVX512-NEXT: vpermq {{.*#+}} zmm18 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpternlogq $226, %zmm18, %zmm21, %zmm17
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm12
+; AVX512-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm8 # 32-byte Folded Reload
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm19
+; AVX512-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm8 = mem[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm8
+; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm15[0,0,1,0]
+; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm14, %zmm11
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm11
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm11
+; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm31[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm0
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm0
; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm14
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm14
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm10
-; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm27[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,0,4,4,5,4]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
-; AVX512-NEXT: # zmm2 = zmm2[2,3,2,3],mem[2,3,2,3]
-; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm15
-; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm15
+; AVX512-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 64-byte Folded Reload
+; AVX512-NEXT: # zmm3 = zmm3[0,1,2,3],mem[2,3,2,3]
+; AVX512-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm9
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm9
+; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm3 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm3
+; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm10[0,1,0,1,4,5,4,5]
+; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,1,0,4,4,5,4]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm5
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm16
+; AVX512-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm16
; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-NEXT: vmovdqa64 %zmm15, 128(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512-NEXT: vmovdqa64 %zmm10, 320(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm14, 256(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm4, 384(%rax)
-; AVX512-NEXT: vmovdqa64 %zmm16, 64(%rax)
-; AVX512-NEXT: addq $1720, %rsp # imm = 0x6B8
+; AVX512-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm0, 256(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm11, 192(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm8, 384(%rax)
+; AVX512-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512-NEXT: addq $1384, %rsp # imm = 0x568
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i8_stride7_vf64:
; AVX512-FCP: # %bb.0:
; AVX512-FCP-NEXT: subq $1432, %rsp # imm = 0x598
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm25
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm25
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128,25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128]
-; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm2[28],zero,ymm2[30,31,30,31],zero,ymm2[29],zero,ymm2[31,28,29]
+; AVX512-FCP-NEXT: 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: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm28
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
+; AVX512-FCP-NEXT: vmovdqa %ymm1, %ymm2
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm15, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm15[30],zero,ymm15[28,u,u,u],zero,ymm15[31],zero,ymm15[29,u]
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm3
; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm23
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm21
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm3[27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm20
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
@@ -9038,10 +9003,10 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm6
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa (%rdi), %ymm5
+; 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: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm3
@@ -9051,366 +9016,351 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[13,u,u,u,u,u],zero,ymm0[14,u,u,u,u,u],zero,ymm0[15,u,u,u,u,u],zero,ymm0[16,u,u,u,u,u],zero,ymm0[17,u,u,u]
-; AVX512-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm14
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm14[13,u,u,u,u,u],zero,ymm14[14,u,u,u,u,u],zero,ymm14[15,u,u,u,u,u],zero,ymm14[16,u,u,u,u,u],zero,ymm14[17,u,u,u]
+; AVX512-FCP-NEXT: vpor %ymm4, %ymm7, %ymm4
+; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm9
; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
-; AVX512-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm7
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vpor %xmm7, %xmm9, %xmm7
-; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpor %xmm7, %xmm9, %xmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm9
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm11, %xmm11
-; AVX512-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm11
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm28
+; AVX512-FCP-NEXT: vpor %xmm9, %xmm11, %xmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm9
; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm14
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm9, %xmm9
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm11
-; AVX512-FCP-NEXT: vmovdqa64 %xmm14, %xmm30
-; AVX512-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm9
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm7
-; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
-; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm21
-; AVX512-FCP-NEXT: vpor %xmm7, %xmm10, %xmm7
-; AVX512-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm7
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm6
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm8
-; AVX512-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
-; AVX512-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm11
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm13
-; AVX512-FCP-NEXT: vmovdqa %xmm6, %xmm9
-; AVX512-FCP-NEXT: vporq %xmm12, %xmm13, %xmm31
-; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm4, %ymm12
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm13
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm6
-; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm12
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm13
-; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm20
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm29
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm12
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
-; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm13
-; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm28
-; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm13
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm12
+; AVX512-FCP-NEXT: vpor %xmm9, %xmm12, %xmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm0, %xmm7
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm24
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm10
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm17
+; AVX512-FCP-NEXT: vpor %xmm7, %xmm10, %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm10, %xmm4
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm8
+; AVX512-FCP-NEXT: vpor %xmm4, %xmm8, %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm11
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm7
+; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm13
+; AVX512-FCP-NEXT: vporq %xmm11, %xmm13, %xmm30
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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: # ymm12 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm15
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm27
-; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm13
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm15
-; AVX512-FCP-NEXT: vpor %ymm13, %ymm15, %ymm6
-; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm14
-; AVX512-FCP-NEXT: vpor %ymm15, %ymm14, %ymm6
-; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
-; AVX512-FCP-NEXT: vpor %ymm5, %ymm4, %ymm4
-; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm25
+; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %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, %ymm1, %ymm11
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm13
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm16
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29],zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm5[28],zero,ymm5[30,31,30,31],zero,ymm5[29],zero,ymm5[31,28,29]
+; AVX512-FCP-NEXT: vporq %ymm6, %ymm5, %ymm20
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm23
+; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27,u,u,u],zero,ymm0[30],zero,ymm0[28,u,u,u],zero,ymm0[31],zero
-; AVX512-FCP-NEXT: vporq %ymm1, %ymm0, %ymm22
-; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,0,1,14],zero,ymm3[14,15,0,1,14,15],zero,ymm3[13,14,15,16,17,16],zero,ymm3[30,31,30,31,16,17],zero,ymm3[31,28,29,30,31]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm14[27,u,u,u],zero,ymm14[30],zero,ymm14[28,u,u,u],zero,ymm14[31],zero
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm0, %ymm21
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero,zero,zero,zero
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm15, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm3
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm29
+; AVX512-FCP-NEXT: vmovdqa64 %ymm23, %ymm2
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm1
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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 {{.*#+}} ymm1 = 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, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm11[8],xmm9[9],xmm11[9],xmm9[10],xmm11[10],xmm9[11],xmm11[11],xmm9[12],xmm11[12],xmm9[13],xmm11[13],xmm9[14],xmm11[14],xmm9[15],xmm11[15]
-; AVX512-FCP-NEXT: vmovdqa %xmm9, %xmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm11, %xmm17
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [2,2,3,3,2,2,3,3]
-; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
-; AVX512-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm12
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm12[13],zero,zero,zero,zero,zero,zero,ymm12[14],zero,zero,zero,zero,zero,zero,ymm12[15],zero,zero,zero,zero,zero,zero,ymm12[16],zero,zero,zero,zero,zero,zero,ymm12[17],zero,zero
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm20
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm1
+; 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: vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5,5,6]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm23
+; AVX512-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm15
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm18
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm27
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512-FCP-NEXT: vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm2[8],xmm11[9],xmm2[9],xmm11[10],xmm2[10],xmm11[11],xmm2[11],xmm11[12],xmm2[12],xmm11[13],xmm2[13],xmm11[14],xmm2[14],xmm11[15],xmm2[15]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm1[8],xmm12[9],xmm1[9],xmm12[10],xmm1[10],xmm12[11],xmm1[11],xmm12[12],xmm1[12],xmm12[13],xmm1[13],xmm12[14],xmm1[14],xmm12[15],xmm1[15]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
+; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm4
+; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[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, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm24
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %xmm28, %xmm5
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm5[8],xmm13[9],xmm5[9],xmm13[10],xmm5[10],xmm13[11],xmm5[11],xmm13[12],xmm5[12],xmm13[13],xmm5[13],xmm13[14],xmm5[14],xmm13[15],xmm5[15]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm18
-; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm21
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm14[8],xmm4[8],xmm14[9],xmm4[9],xmm14[10],xmm4[10],xmm14[11],xmm4[11],xmm14[12],xmm4[12],xmm14[13],xmm4[13],xmm14[14],xmm4[14],xmm14[15],xmm4[15]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm13
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm5
-; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm15[8],xmm5[9],xmm15[9],xmm5[10],xmm15[10],xmm5[11],xmm15[11],xmm5[12],xmm15[12],xmm5[13],xmm15[13],xmm5[14],xmm15[14],xmm5[15],xmm15[15]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm17
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm6
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm14 = xmm6[0,1,2,3,4,5,5,6]
+; AVX512-FCP-NEXT: vpermd %ymm14, %ymm3, %ymm31
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm8[8],xmm7[9],xmm8[9],xmm7[10],xmm8[10],xmm7[11],xmm8[11],xmm7[12],xmm8[12],xmm7[13],xmm8[13],xmm7[14],xmm8[14],xmm7[15],xmm8[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm3, %xmm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm3
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,6]
-; AVX512-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm26
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm24 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm11, %zmm3, %zmm28
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
+; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm26 # 16-byte Folded Reload
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm9
; AVX512-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm2
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm8, %xmm8
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm31[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm14, %xmm1
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm30 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm31, %zmm30, %zmm30
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm30
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm31 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512-FCP-NEXT: vpternlogq $184, %zmm30, %zmm31, %zmm14
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm18[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm13
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm7, %zmm29, %zmm7
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm27 = zmm27[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vporq %zmm28, %zmm27, %zmm27
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm27
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm23[2,3,2,3],zmm7[0,1,0,1]
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3],xmm5[4],xmm15[4],xmm5[5],xmm15[5],xmm5[6],xmm15[6],xmm5[7],xmm15[7]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm10
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[2,3,2,3],zmm10[0,1,0,1]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm19, %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,0,1,2,0,0,1]
-; AVX512-FCP-NEXT: vpermd %ymm11, %ymm17, %ymm28
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,1,0,0,4,5,6,7]
-; AVX512-FCP-NEXT: vpermd %ymm4, %ymm17, %ymm17
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [4,5,4,5,5,7,4,5]
-; AVX512-FCP-NEXT: vpermd %ymm2, %ymm23, %ymm2
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512-FCP-NEXT: vpermd %ymm15, %ymm23, %ymm15
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm15
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm7
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm1 = mem[2,3,2,3]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm30, %zmm7, %zmm7
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm5[0],xmm13[0],xmm5[1],xmm13[1],xmm5[2],xmm13[2],xmm5[3],xmm13[3],xmm5[4],xmm13[4],xmm5[5],xmm13[5],xmm5[6],xmm13[6],xmm5[7],xmm13[7]
+; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm20[2,3,2,3],zmm8[0,1,0,1]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm20 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm20
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm30 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512-FCP-NEXT: vpternlogq $184, %zmm20, %zmm8, %zmm30
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm19[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm17[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm20
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm5 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm25 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm5[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm25, %zmm8, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm4
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3],xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm5
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm18 = mem[2,3,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm3, %zmm18
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm25[0,1,0,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,0]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,0]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm18
-; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm14 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm14
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm20
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm0
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm24[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm9
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm5
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm5
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm6 = zmm3[2,3,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm4
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm3[2,3,2,3],zmm5[0,1,0,1]
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm5, %xmm5
+; AVX512-FCP-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-FCP-NEXT: vpternlogq $226, %zmm14, %zmm8, %zmm10
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[2,3,2,3],zmm5[0,1,0,1]
+; 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), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm29[2,3,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm14
+; AVX512-FCP-NEXT: vpternlogq $226, %zmm11, %zmm8, %zmm14
+; AVX512-FCP-NEXT: vmovdqa64 %xmm23, %xmm0
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm0[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,0,1,2,0,0,1]
+; AVX512-FCP-NEXT: vpermd %ymm8, %ymm11, %ymm16
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm12
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm5
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[1,1,0,0,4,5,6,7]
+; AVX512-FCP-NEXT: vpermd %ymm0, %ymm11, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = 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: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [4,5,4,5,5,7,4,5]
+; AVX512-FCP-NEXT: vpermd %ymm15, %ymm17, %ymm15
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,ymm13[13],zero,zero,zero,zero,zero,zero,ymm13[14],zero,zero,zero,zero,zero,zero,ymm13[15],zero,zero,zero,zero,zero,zero,ymm13[16],zero,zero,zero,zero,zero,zero,ymm13[17],zero,zero
+; 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, %ymm17, %ymm13
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm24[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm27[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm22[0,1,0,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,0]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm12[0,0,1,0]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm17
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm18
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm18
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm5, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm28[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm5
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm26[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm9[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm16, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm7[0,1,0,1,4,5,4,5]
+; AVX512-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm6 = zmm3[0,1,2,3],mem[2,3,2,3]
-; AVX512-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
-; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm11
+; AVX512-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm6
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm6
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm0
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm0
+; 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 $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm8
+; AVX512-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm8
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 320(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 64(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax)
; AVX512-FCP-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i8_stride7_vf64:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: subq $1720, %rsp # imm = 0x6B8
+; AVX512DQ-NEXT: subq $1384, %rsp # imm = 0x568
; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm7, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm7[14],zero,zero,zero,zero,zero,zero,ymm7[15],zero,zero,zero,zero,zero,zero,ymm7[16],zero,zero,zero,zero,zero,zero,ymm7[17],zero,zero,zero,zero,zero,zero,ymm7[18]
; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm25
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm20
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm14
+; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm15
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm15, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm23
+; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm2
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm8, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm23
-; AVX512DQ-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r8), %ymm15
+; 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, %ymm15, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm24
-; AVX512DQ-NEXT: vmovdqa (%r9), %ymm2
+; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512DQ-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%r9), %ymm8
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0,13,0,0,0,128,16,128,14,0,0,0,128,17,128,15,0]
; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm27
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm8, %ymm1
+; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm17
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm10
@@ -9423,7 +9373,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,14,128,12,128,14,15,14,15,128,13,128,15,12,13,28,29,30,128,28,128,30,31,30,31,128,29,128,31,28,29]
; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm16
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm19
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm10[23,u,u,u],zero,ymm10[26],zero,ymm10[24,u,u,u],zero,ymm10[27],zero
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -9438,7 +9388,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0,27,0,0,0,128,30,128,28,0,0,0,128,31,128,29,0]
; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm21
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm5[23],zero,ymm5[21,22,23,26],zero,ymm5[24],zero,ymm5[28,29,26,27]
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
@@ -9446,7 +9396,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29,128,27,128,128,128,128,30,128,28,128,128,128,128,31,128,29]
; AVX512DQ-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm11
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm21
+; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm25
; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm2
; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
@@ -9465,23 +9415,23 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vpshufb %ymm6, %ymm7, %ymm6
-; AVX512DQ-NEXT: vmovdqa64 %ymm7, %ymm17
-; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm13
-; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm13, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm7, %ymm22
+; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm7, %ymm7
; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm8, %ymm6
-; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm14, %ymm7
-; AVX512DQ-NEXT: vmovdqa64 %ymm14, %ymm22
+; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm12
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm12, %ymm6
+; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm15, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm15, %ymm19
; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm15, %ymm6
-; AVX512DQ-NEXT: vmovdqa64 %ymm15, %ymm16
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm15
-; AVX512DQ-NEXT: vpshufb %ymm11, %ymm15, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512DQ-NEXT: vpshufb %ymm11, %ymm8, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm8, %ymm16
; AVX512DQ-NEXT: vpor %ymm6, %ymm7, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm0
@@ -9490,379 +9440,348 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
; AVX512DQ-NEXT: vpshufb %xmm11, %xmm1, %xmm6
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm12 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm7
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm7
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm21
; AVX512DQ-NEXT: vpor %xmm6, %xmm7, %xmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm15
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm6 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
-; AVX512DQ-NEXT: vpshufb %xmm6, %xmm14, %xmm7
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm8
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm18
+; AVX512DQ-NEXT: vpshufb %xmm6, %xmm15, %xmm7
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm13 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm0, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm30
; AVX512DQ-NEXT: vpor %xmm7, %xmm8, %xmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,2,3,0,1,0,18,0,19,18,0,19,0]
; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm0
; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm0[0,1,2,3,4,5,5,6]
; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm8 = xmm0[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm8, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vpermi2d %zmm7, %zmm8, %zmm28
; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm1
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-NEXT: vpshufb %xmm8, %xmm0, %xmm7
-; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm20
+; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm7
+; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm24
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm18
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm30
-; AVX512DQ-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512DQ-NEXT: vpor %xmm7, %xmm8, %xmm0
-; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm28, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm7
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm14, %xmm8
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm29
+; AVX512DQ-NEXT: vporq %xmm7, %xmm8, %xmm31
; AVX512DQ-NEXT: vmovdqa64 %ymm23, %ymm0
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm7
+; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm0
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm8
; AVX512DQ-NEXT: vpor %ymm7, %ymm8, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
; AVX512DQ-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512DQ-NEXT: vpshufb %ymm1, %ymm4, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
-; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm1
-; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm19
+; AVX512DQ-NEXT: vmovdqa64 %ymm1, %ymm23
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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: # ymm7 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm7, %ymm5, %ymm1
; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm26, %ymm0
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm10, %ymm0
-; AVX512DQ-NEXT: vmovdqa64 %ymm25, %ymm1
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm1
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm10[14],zero,zero,zero,zero,zero,zero,ymm10[15],zero,zero,zero,zero,zero,zero,ymm10[16],zero,zero,zero,zero,zero,zero,ymm10[17],zero,zero,zero,zero,zero,zero,ymm10[18]
+; AVX512DQ-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-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX512DQ-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm10, %ymm0
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm1 = [2,3,4,5,128,3,128,5,4,5,6,128,4,128,6,7,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm9, %ymm4
-; AVX512DQ-NEXT: vpor %ymm0, %ymm4, %ymm0
-; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512DQ-NEXT: # ymm10 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm2, %ymm4
-; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512DQ-NEXT: 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-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm5
+; AVX512DQ-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm8 = [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 %ymm8, %ymm9, %ymm4
+; AVX512DQ-NEXT: vpor %ymm1, %ymm4, %ymm1
+; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-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, %ymm4
+; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [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: # ymm9 = mem[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm3, %ymm5
; AVX512DQ-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm24, %ymm4
-; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm3
; AVX512DQ-NEXT: vmovdqa64 %ymm27, %ymm4
+; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm4
; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-NEXT: vpor %ymm3, %ymm2, %ymm2
; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm3
; AVX512DQ-NEXT: vpshufb %xmm6, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm31
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm25
; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm3
-; AVX512DQ-NEXT: vpshufb %xmm3, %xmm4, %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm28
+; AVX512DQ-NEXT: vpshufb %xmm13, %xmm4, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm27
; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm3
; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm6
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512DQ-NEXT: vpshufb %xmm12, %xmm4, %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm18
+; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm17
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm13
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm13, %xmm3
; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm2
; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%r9), %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm3, %xmm2
-; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm8
-; AVX512DQ-NEXT: vmovdqa (%r8), %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm4
-; AVX512DQ-NEXT: vpshufb %xmm4, %xmm3, %xmm4
-; AVX512DQ-NEXT: vmovdqa %xmm3, %xmm9
-; AVX512DQ-NEXT: vpor %xmm2, %xmm4, %xmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[23],zero,zero,zero,zero,ymm13[26],zero,ymm13[24],zero,zero,zero,zero,ymm13[27],zero,ymm13[25]
-; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm7, %ymm3, %ymm4
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25],zero,ymm3[23,u,u,u],zero,ymm3[26],zero,ymm3[24,u,u,u],zero,ymm3[27],zero
-; AVX512DQ-NEXT: vpshufb %ymm1, %ymm13, %ymm1
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm21, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
-; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm3
-; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm15[25],zero,ymm15[23],zero,zero,zero,zero,ymm15[26],zero,ymm15[24],zero,zero
-; AVX512DQ-NEXT: vmovdqa64 %ymm16, %ymm3
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm3[23],zero,ymm3[23,24,25,26],zero,ymm3[24],zero,ymm3[30,31]
-; AVX512DQ-NEXT: vpor %ymm1, %ymm2, %ymm1
-; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vpshufb %ymm10, %ymm15, %ymm1
+; AVX512DQ-NEXT: vmovdqa (%r9), %xmm11
+; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm2
+; AVX512DQ-NEXT: vpshufb %xmm2, %xmm11, %xmm2
+; AVX512DQ-NEXT: vmovdqa (%r8), %xmm10
+; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-NEXT: vpshufb %xmm3, %xmm10, %xmm3
+; AVX512DQ-NEXT: vpor %xmm2, %xmm3, %xmm2
+; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-NEXT: vmovdqa64 %ymm20, %ymm4
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[23],zero,zero,zero,zero,ymm4[26],zero,ymm4[24],zero,zero,zero,zero,ymm4[27],zero,ymm4[25]
+; AVX512DQ-NEXT: vmovdqa64 %ymm22, %ymm3
; AVX512DQ-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-NEXT: vmovdqa (%rax), %ymm4
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm15
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm9, %xmm21
-; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm22
-; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm24
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm1[0,1,0,1],zmm0[4,5,6,7]
-; AVX512DQ-NEXT: vmovdqa (%rax), %xmm11
-; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,5,5,6]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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 %ymm8, %ymm4, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm29
+; AVX512DQ-NEXT: vmovdqa64 %ymm19, %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 %ymm23, %ymm3
+; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm26
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = 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: vpshufb %ymm7, %ymm12, %ymm2
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm23
+; 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,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
+; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; 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 %ymm9, %ymm1, %ymm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-NEXT: vmovdqa (%rax), %ymm8
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,5,4,0,5,0,4,0,20,21,0,23,0,21,0,23]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm2 = ymm8[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm16
+; AVX512DQ-NEXT: vmovdqa (%rax), %xmm5
+; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm5[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 {{.*#+}} ymm9 = [128,13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128]
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm4, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3],xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm8, %ymm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm19
+; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm4
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm4[0],xmm15[0],xmm4[1],xmm15[1],xmm4[2],xmm15[2],xmm4[3],xmm15[3],xmm4[4],xmm15[4],xmm4[5],xmm15[5],xmm4[6],xmm15[6],xmm4[7],xmm15[7]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm2, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm27
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm24
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[2,3,2,3],zmm2[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm0[2,3,2,3],zmm3[0,1,0,1]
; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
; AVX512DQ-NEXT: vpshufb %xmm0, %xmm3, %xmm3
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm26
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm21
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[2,3,2,3],zmm3[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm3
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm0[2,3,2,3],zmm3[0,1,0,1]
+; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm2
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm14[0],xmm2[0],xmm14[1],xmm2[1],xmm14[2],xmm2[2],xmm14[3],xmm2[3],xmm14[4],xmm2[4],xmm14[5],xmm2[5],xmm14[6],xmm2[6],xmm14[7],xmm2[7]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm25
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm12, %xmm12
+; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm20
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[2,3,2,3],zmm13[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512DQ-NEXT: vmovdqa %xmm6, %xmm8
-; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm7
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
+; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[2,3,2,3],zmm12[0,1,0,1]
+; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm8[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, %ymm18
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm7[8],xmm1[8],xmm7[9],xmm1[9],xmm7[10],xmm1[10],xmm7[11],xmm1[11],xmm7[12],xmm1[12],xmm7[13],xmm1[13],xmm7[14],xmm1[14],xmm7[15],xmm1[15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm17, %xmm3
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm3[8],xmm13[8],xmm3[9],xmm13[9],xmm3[10],xmm13[10],xmm3[11],xmm13[11],xmm3[12],xmm13[12],xmm3[13],xmm13[13],xmm3[14],xmm13[14],xmm3[15],xmm13[15]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm4, %xmm2
-; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-NEXT: vpshufb %xmm1, %xmm8, %xmm12
; AVX512DQ-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm6
-; AVX512DQ-NEXT: vmovdqa64 %xmm28, %xmm4
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm8
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm4[8],xmm15[9],xmm4[9],xmm15[10],xmm4[10],xmm15[11],xmm4[11],xmm15[12],xmm4[12],xmm15[13],xmm4[13],xmm15[14],xmm4[14],xmm15[15],xmm4[15]
+; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm4
+; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm7
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm4[8],xmm7[8],xmm4[9],xmm7[9],xmm4[10],xmm7[10],xmm4[11],xmm7[11],xmm4[12],xmm7[12],xmm4[13],xmm7[13],xmm4[14],xmm7[14],xmm4[15],xmm7[15]
; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-NEXT: vpshufb %xmm0, %xmm14, %xmm13
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm15, %xmm15
; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm0
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
-; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15]
-; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm2[0,1,0,1],zmm1[0,1,0,1]
-; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-NEXT: vpshufb %ymm9, %ymm10, %ymm1
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm17
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm2[8],xmm14[9],xmm2[9],xmm14[10],xmm2[10],xmm14[11],xmm2[11],xmm14[12],xmm2[12],xmm14[13],xmm2[13],xmm14[14],xmm2[14],xmm14[15],xmm2[15]
+; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm10[8],xmm11[8],xmm10[9],xmm11[9],xmm10[10],xmm11[10],xmm10[11],xmm11[11],xmm10[12],xmm11[12],xmm10[13],xmm11[13],xmm10[14],xmm11[14],xmm10[15],xmm11[15]
+; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm14, %xmm14
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm31, %zmm31
+; AVX512DQ-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-NEXT: vpshufb %ymm9, %ymm0, %ymm1
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm9 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm12, %zmm10, %zmm17
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm2 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm28 # 32-byte Folded Reload
-; AVX512DQ-NEXT: # ymm28 = mem[2,3,2,3]
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm28, %zmm3, %zmm28
-; AVX512DQ-NEXT: vpternlogq $226, %zmm2, %zmm10, %zmm28
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm9[2,2,3,2]
-; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm10
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm2
-; AVX512DQ-NEXT: vpshufb %xmm2, %xmm1, %xmm1
-; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm27 # 16-byte Folded Reload
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm2, %xmm2
-; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 16-byte Folded Reload
-; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm1
-; AVX512DQ-NEXT: vmovdqa64 %xmm22, %xmm3
-; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm1
-; AVX512DQ-NEXT: vpshufb %xmm1, %xmm8, %xmm8
-; AVX512DQ-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm8 = zmm8[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm9 = xmm11[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm11[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,2]
+; AVX512DQ-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm9
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm9
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm1 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm6, %zmm1, %zmm22
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm6 = mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
+; AVX512DQ-NEXT: # ymm27 = mem[2,3,2,3]
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm27, %zmm0, %zmm27
+; AVX512DQ-NEXT: vpternlogq $226, %zmm6, %zmm1, %zmm27
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3],xmm7[4],xmm4[4],xmm7[5],xmm4[5],xmm7[6],xmm4[6],xmm7[7],xmm4[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm1, %xmm1
+; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 16-byte Folded Reload
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm13[0],xmm3[0],xmm13[1],xmm3[1],xmm13[2],xmm3[2],xmm13[3],xmm3[3],xmm13[4],xmm3[4],xmm13[5],xmm3[5],xmm13[6],xmm3[6],xmm13[7],xmm3[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm6, %xmm6
+; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 16-byte Folded Reload
+; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3],xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
+; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm0
+; AVX512DQ-NEXT: vpshufb %xmm0, %xmm10, %xmm10
+; AVX512DQ-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 16-byte Folded Reload
+; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm5[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[1,1,0,0,4,5,6,7]
; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,2,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm9, %zmm11, %zmm9
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm11 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm11, %zmm12, %zmm11
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm11
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm9 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm12 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm9, %zmm12, %zmm9
-; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm12 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512DQ-NEXT: vpternlogq $184, %zmm11, %zmm12, %zmm9
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm13[0,1,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm18[0,1,0,1]
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm20[2,3,2,3]
-; AVX512DQ-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; AVX512DQ-NEXT: # xmm1 = mem[1,1,0,0,4,5,6,7]
-; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,0]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm13 = zmm19[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm0
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm21 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm13, %zmm21, %zmm13
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm20 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm18 = zmm30[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vporq %zmm20, %zmm18, %zmm18
-; AVX512DQ-NEXT: vpternlogq $226, %zmm13, %zmm12, %zmm18
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm4
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm23, %zmm16
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm16
-; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm4 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
-; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,0]
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm1
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm1
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm5
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm13 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm13 = zmm13[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm13
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm21 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512DQ-NEXT: vpternlogq $184, %zmm13, %zmm21, %zmm11
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm17[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm8, %zmm21, %zmm7
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm15[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm14[0,1,0,1]
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,3,2,3]
+; AVX512DQ-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512DQ-NEXT: # xmm15 = mem[1,1,0,0,4,5,6,7]
+; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
+; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm20 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vporq %zmm26, %zmm23, %zmm17
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm18 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm17 = zmm17[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpternlogq $226, %zmm18, %zmm21, %zmm17
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm12
+; AVX512DQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm19
+; AVX512DQ-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm8 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm8
+; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm15[0,0,1,0]
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm14, %zmm11
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm11
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm11
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm31[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm0
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm0
; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm14
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm14
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm0 = zmm0[0,1,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm10
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm27[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
-; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm5[0,0,1,0,4,4,5,4]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm0
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-NEXT: # zmm2 = zmm2[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm15
-; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm15
+; AVX512DQ-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-NEXT: # zmm3 = zmm3[0,1,2,3],mem[2,3,2,3]
+; AVX512DQ-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm9
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm9
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm3 = zmm6[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm3
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm10[0,1,0,1,4,5,4,5]
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,1,0,4,4,5,4]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm5
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm16
+; AVX512DQ-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm16
; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-NEXT: vmovdqa64 %zmm15, 128(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm10, 320(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm14, 256(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm4, 384(%rax)
-; AVX512DQ-NEXT: vmovdqa64 %zmm16, 64(%rax)
-; AVX512DQ-NEXT: addq $1720, %rsp # imm = 0x6B8
+; AVX512DQ-NEXT: vmovdqa64 %zmm16, 128(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm0, 256(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm11, 192(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm8, 384(%rax)
+; AVX512DQ-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512DQ-NEXT: addq $1384, %rsp # imm = 0x568
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i8_stride7_vf64:
; AVX512DQ-FCP: # %bb.0:
; AVX512DQ-FCP-NEXT: subq $1432, %rsp # imm = 0x598
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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-FCP-NEXT: vmovdqa 32(%rdi), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm25
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm3[28],zero,ymm3[30,31,30,31],zero,ymm3[29],zero,ymm3[31,28,29]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm25
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128,25,128,23,0,0,0,128,26,128,24,0,0,0,128,27,128]
-; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm2[28],zero,ymm2[30,31,30,31],zero,ymm2[29],zero,ymm2[31,28,29]
+; AVX512DQ-FCP-NEXT: 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: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm27
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm18
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
-; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm28
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm15
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm15, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm29
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm19
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm15[30],zero,ymm15[28,u,u,u],zero,ymm15[31],zero,ymm15[29,u]
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27,24,25,128,23,128,21,22,23,26,128,24,128,28,29,26,27]
; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm24
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm31
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[27],zero,zero,zero,zero,ymm3[30],zero,ymm3[28],zero,zero,zero,zero,ymm3[31],zero,ymm3[29]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[27],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm3
; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128]
; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm23
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm21
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm2[27,u,u,u],zero,ymm2[30],zero,ymm2[28,u,u,u],zero,ymm2[31],zero
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31,24,25,24,25,128,23,128,23,24,25,26,128,24,128,30,31]
-; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm3[27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm20
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
@@ -9872,10 +9791,10 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm5[14],zero,zero,zero,zero,zero,zero,ymm5[15],zero,zero,zero,zero,zero,zero,ymm5[16],zero,zero,zero,zero,zero,zero,ymm5[17],zero,zero,zero,zero,zero,zero,ymm5[18]
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm4[0,1,14],zero,ymm4[12,13,0,1,14,15],zero,ymm4[3,12,13,2,3,16],zero,ymm4[30,31,28,29,16,17],zero,ymm4[31,18,19,28,29,18],zero
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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: vmovdqa (%rdi), %ymm5
+; 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: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm3
@@ -9885,327 +9804,316 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[13,u,u,u,u,u],zero,ymm0[14,u,u,u,u,u],zero,ymm0[15,u,u,u,u,u],zero,ymm0[16,u,u,u,u,u],zero,ymm0[17,u,u,u]
-; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm14
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm14[13,u,u,u,u,u],zero,ymm14[14,u,u,u,u,u],zero,ymm14[15,u,u,u,u,u],zero,ymm14[16,u,u,u,u,u],zero,ymm14[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm7, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm9
; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm0, %xmm7
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u]
; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm9, %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm9, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm0
; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128]
; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm9
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm11, %xmm11
-; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm28
+; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm11, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm9
; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm14
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm9, %xmm9
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm14, %xmm30
-; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm11, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm9
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm21
-; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm10, %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm7
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm7, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm8
-; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm8, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm11
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm13
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, %xmm9
-; AVX512DQ-FCP-NEXT: vporq %xmm12, %xmm13, %xmm31
-; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm4, %ymm12
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,128]
-; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm5, %ymm13
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm12
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23,18,19,20,21,128,19,128,21,20,21,22,128,20,128,22,23]
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm20
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm29
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm12
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20]
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm16
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm28
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm13
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm12
+; AVX512DQ-FCP-NEXT: vpor %xmm9, %xmm12, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm0, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm24
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm0, %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm17
+; AVX512DQ-FCP-NEXT: vpor %xmm7, %xmm10, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm10, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm9, %xmm8
+; AVX512DQ-FCP-NEXT: vpor %xmm4, %xmm8, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm7
+; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm13
+; AVX512DQ-FCP-NEXT: vporq %xmm11, %xmm13, %xmm30
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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: # ymm12 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm15
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm27
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm13
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm15
-; AVX512DQ-FCP-NEXT: vpor %ymm13, %ymm15, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm24
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
-; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm14
-; AVX512DQ-FCP-NEXT: vpor %ymm15, %ymm14, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29],zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm4[28],zero,ymm4[30,31,30,31],zero,ymm4[29],zero,ymm4[31,28,29]
-; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128,18,128,18,19,20,21,128,19,128,25,26,27,22,128,20,128]
+; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm25
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm11
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm16
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29],zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,30],zero,ymm5[28],zero,ymm5[30,31,30,31],zero,ymm5[29],zero,ymm5[31,28,29]
+; AVX512DQ-FCP-NEXT: vporq %ymm6, %ymm5, %ymm20
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[30],zero,ymm2[28],zero,zero,zero,zero,ymm2[31],zero,ymm2[29],zero,zero
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u],zero,ymm3[30],zero,ymm3[28,u,u,u],zero,ymm3[31],zero,ymm3[29,u]
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm23
+; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[27],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm0[27,u,u,u],zero,ymm0[30],zero,ymm0[28,u,u,u],zero,ymm0[31],zero
-; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm0, %ymm22
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[0,1,0,1,14],zero,ymm3[14,15,0,1,14,15],zero,ymm3[13,14,15,16,17,16],zero,ymm3[30,31,30,31,16,17],zero,ymm3[31,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm14[27,u,u,u],zero,ymm14[30],zero,ymm14[28,u,u,u],zero,ymm14[31],zero
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm0, %ymm21
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,0,1,14],zero,ymm2[14,15,0,1,14,15],zero,ymm2[13,14,15,16,17,16],zero,ymm2[30,31,30,31,16,17],zero,ymm2[31,28,29,30,31]
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm15, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm3
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm3[14],zero,zero,zero,zero,zero,zero,ymm3[15],zero,zero,zero,zero,zero,zero,ymm3[16],zero,zero,zero,zero,zero,zero,ymm3[17],zero,zero,zero,zero,zero,zero,ymm3[18]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[12,13,0,1,14,15],zero,ymm2[3,12,13,2,3,16],zero,ymm2[30,31,28,29,16,17],zero,ymm2[31,18,19,28,29,18],zero
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm29
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm1
; 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 = zero,zero,zero,zero,zero,zero,ymm2[14],zero,zero,zero,zero,zero,zero,ymm2[15],zero,zero,zero,zero,zero,zero,ymm2[16],zero,zero,zero,zero,zero,zero,ymm2[17],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[13,u,u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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 {{.*#+}} ymm1 = 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, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm11[8],xmm9[9],xmm11[9],xmm9[10],xmm11[10],xmm9[11],xmm11[11],xmm9[12],xmm11[12],xmm9[13],xmm11[13],xmm9[14],xmm11[14],xmm9[15],xmm11[15]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm11, %xmm17
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [2,2,3,3,2,2,3,3]
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm0
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,6]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm19
-; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = zero,ymm12[13],zero,zero,zero,zero,zero,zero,ymm12[14],zero,zero,zero,zero,zero,zero,ymm12[15],zero,zero,zero,zero,zero,zero,ymm12[16],zero,zero,zero,zero,zero,zero,ymm12[17],zero,zero
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm20
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm1
+; 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: vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5,5,6]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm23
+; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm18
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm27
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512DQ-FCP-NEXT: vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm11[8],xmm2[8],xmm11[9],xmm2[9],xmm11[10],xmm2[10],xmm11[11],xmm2[11],xmm11[12],xmm2[12],xmm11[13],xmm2[13],xmm11[14],xmm2[14],xmm11[15],xmm2[15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm1
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm1[8],xmm12[9],xmm1[9],xmm12[10],xmm1[10],xmm12[11],xmm1[11],xmm12[12],xmm1[12],xmm12[13],xmm1[13],xmm12[14],xmm1[14],xmm12[15],xmm1[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm19
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm1
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[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, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm24
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm5
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm5[8],xmm13[9],xmm5[9],xmm13[10],xmm5[10],xmm13[11],xmm5[11],xmm13[12],xmm5[12],xmm13[13],xmm5[13],xmm13[14],xmm5[14],xmm13[15],xmm5[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm18
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm0
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm21
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm14[8],xmm4[8],xmm14[9],xmm4[9],xmm14[10],xmm4[10],xmm14[11],xmm4[11],xmm14[12],xmm4[12],xmm14[13],xmm4[13],xmm14[14],xmm4[14],xmm14[15],xmm4[15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm13
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm5
-; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm5[8],xmm15[8],xmm5[9],xmm15[9],xmm5[10],xmm15[10],xmm5[11],xmm15[11],xmm5[12],xmm15[12],xmm5[13],xmm15[13],xmm5[14],xmm15[14],xmm5[15],xmm15[15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm17
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm6
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm14 = xmm6[0,1,2,3,4,5,5,6]
+; AVX512DQ-FCP-NEXT: vpermd %ymm14, %ymm3, %ymm31
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm8[8],xmm7[9],xmm8[9],xmm7[10],xmm8[10],xmm7[11],xmm8[11],xmm7[12],xmm8[12],xmm7[13],xmm8[13],xmm7[14],xmm8[14],xmm7[15],xmm8[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm3, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm3
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,6]
-; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm6, %ymm26
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm24 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm11, %zmm3, %zmm28
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm26 # 16-byte Folded Reload
+; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm9
; AVX512DQ-FCP-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm8, %xmm8
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm8[0,1,0,1],zmm31[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm14, %xmm1
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm30 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm31, %zmm30, %zmm30
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm30
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm14 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm31 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm14, %zmm31, %zmm14
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm31 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm30, %zmm31, %zmm14
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm18[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm13 = zmm13[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm13
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm7, %zmm29, %zmm7
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm27 = zmm27[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vporq %zmm28, %zmm27, %zmm27
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm7, %zmm31, %zmm27
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[2,3,2,3],zmm1[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm23[2,3,2,3],zmm7[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3],xmm5[4],xmm15[4],xmm5[5],xmm15[5],xmm5[6],xmm15[6],xmm5[7],xmm15[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm10, %xmm10
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[2,3,2,3],zmm10[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm2
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [0,1,0,1,2,0,0,1]
-; AVX512DQ-FCP-NEXT: vpermd %ymm11, %ymm17, %ymm28
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,1,0,0,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm17, %ymm17
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm10 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm2 = ymm12[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm23 = [4,5,4,5,5,7,4,5]
-; AVX512DQ-FCP-NEXT: vpermd %ymm2, %ymm23, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = zero,ymm15[13],zero,zero,zero,zero,zero,zero,ymm15[14],zero,zero,zero,zero,zero,zero,ymm15[15],zero,zero,zero,zero,zero,zero,ymm15[16],zero,zero,zero,zero,zero,zero,ymm15[17],zero,zero
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpermd %ymm15, %ymm23, %ymm15
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm15, %ymm15
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm7
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm1 = mem[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm30, %zmm7, %zmm7
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm5[0],xmm13[0],xmm5[1],xmm13[1],xmm5[2],xmm13[2],xmm5[3],xmm13[3],xmm5[4],xmm13[4],xmm5[5],xmm13[5],xmm5[6],xmm13[6],xmm5[7],xmm13[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm20[2,3,2,3],zmm8[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm20 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm30 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255]
+; AVX512DQ-FCP-NEXT: vpternlogq $184, %zmm20, %zmm8, %zmm30
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm19[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm20 = zmm17[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm5, %zmm8, %zmm20
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm25, %zmm5 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm25 = zmm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm5[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm25, %zmm8, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm4
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3],xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm5
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm18 = mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm3, %zmm18
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm21[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm25[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,0]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,0]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm1, %zmm15, %zmm18
-; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm14 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm14
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm20
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm20
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm26, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm0
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm24[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,1,0,1,4,5,4,5]
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm9
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm28, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $85, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm6 = zmm3[2,3,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm22, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm3[2,3,2,3],zmm5[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm5, %xmm5
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq $226, %zmm14, %zmm8, %zmm10
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[2,3,2,3],zmm5[0,1,0,1]
+; 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), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm29[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm14
+; AVX512DQ-FCP-NEXT: vpternlogq $226, %zmm11, %zmm8, %zmm14
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm23, %xmm0
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm0[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,0,1,2,0,0,1]
+; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm11, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm5
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[1,1,0,0,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm11, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = 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: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm17 = [4,5,4,5,5,7,4,5]
+; AVX512DQ-FCP-NEXT: vpermd %ymm15, %ymm17, %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = zero,ymm13[13],zero,zero,zero,zero,zero,zero,ymm13[14],zero,zero,zero,zero,zero,zero,ymm13[15],zero,zero,zero,zero,zero,zero,ymm13[16],zero,zero,zero,zero,zero,zero,ymm13[17],zero,zero
+; 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, %ymm17, %ymm13
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm24[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm27[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm22[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,0]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm12[0,0,1,0]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm30, %zmm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm17
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm18
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm17, %zmm18
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm31, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm28[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm5
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm26[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm9[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm16, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm4 = zmm7[0,1,0,1,4,5,4,5]
+; AVX512DQ-FCP-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vshufi64x2 $84, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm6 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm6 = zmm3[0,1,2,3],mem[2,3,2,3]
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
-; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm11
+; AVX512DQ-FCP-NEXT: vporq {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm6
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm3[2,3,2,3,6,7,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm6
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm6
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm0
+; 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 $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm8
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 320(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 64(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 384(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax)
; AVX512DQ-FCP-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
index 1771b53a0f835..f5a6b9f59aacf 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
@@ -2104,16 +2104,14 @@ define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: kmovd %ecx, %k3
; AVX512BW-NEXT: vmovdqa32 %zmm4, %zmm5 {%k3}
; AVX512BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
-; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,2,10,2,10,u,u,u,u,3,11,3,11,u,u,u,u,4,12,4,12,u,u,u,u,5,13,5,13,u,u,u,u,6,14,6,14,u,u,u,u,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %zmm4, %zmm2, %zmm2
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} zmm3 = zmm3[1,3,1,3,5,7,5,7]
-; AVX512BW-NEXT: vpshufb %zmm4, %zmm3, %zmm3
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63]
; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm2 {%k1}
; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
-; AVX512BW-NEXT: vpmovsxdq {{.*#+}} zmm3 = [134219776,151062785,167905794,184748803,201591812,218434821,235277830,252120839]
-; AVX512BW-NEXT: vpshufb %zmm3, %zmm0, %zmm0
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[1,3,1,3,5,7,5,7]
-; AVX512BW-NEXT: vpshufb %zmm3, %zmm1, %zmm1
+; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u]
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512BW-NEXT: vmovdqa64 %zmm0, 64(%rax)
@@ -2159,16 +2157,14 @@ define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm1 = [5,7,5,7,5,7,5,7]
; AVX512BW-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512BW-FCP-NEXT: vpermq %zmm5, %zmm1, %zmm5
-; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,2,10,2,10,u,u,u,u,3,11,3,11,u,u,u,u,4,12,4,12,u,u,u,u,5,13,5,13,u,u,u,u,6,14,6,14,u,u,u,u,7,15,7,15]
-; AVX512BW-FCP-NEXT: vpshufb %zmm6, %zmm5, %zmm5
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
-; AVX512BW-FCP-NEXT: vpshufb %zmm6, %zmm2, %zmm2
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u]
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm2 {%k1}
; AVX512BW-FCP-NEXT: vpermq %zmm4, %zmm1, %zmm1
-; AVX512BW-FCP-NEXT: vpmovsxdq {{.*#+}} zmm4 = [134219776,151062785,167905794,184748803,201591812,218434821,235277830,252120839]
-; AVX512BW-FCP-NEXT: vpshufb %zmm4, %zmm1, %zmm1
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u]
; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
-; AVX512BW-FCP-NEXT: vpshufb %zmm4, %zmm0, %zmm0
+; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
@@ -2213,16 +2209,14 @@ define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: kmovd %ecx, %k3
; AVX512DQ-BW-NEXT: vmovdqa32 %zmm4, %zmm5 {%k3}
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm4 = [u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,2,10,2,10,u,u,u,u,3,11,3,11,u,u,u,u,4,12,4,12,u,u,u,u,5,13,5,13,u,u,u,u,6,14,6,14,u,u,u,u,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %zmm4, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm3 = zmm3[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-NEXT: vpshufb %zmm4, %zmm3, %zmm3
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63]
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm3, %zmm2 {%k1}
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-NEXT: vpmovsxdq {{.*#+}} zmm3 = [134219776,151062785,167905794,184748803,201591812,218434821,235277830,252120839]
-; AVX512DQ-BW-NEXT: vpshufb %zmm3, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-NEXT: vpshufb %zmm3, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u]
; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 64(%rax)
@@ -2268,16 +2262,14 @@ define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm1 = [5,7,5,7,5,7,5,7]
; AVX512DQ-BW-FCP-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
; AVX512DQ-BW-FCP-NEXT: vpermq %zmm5, %zmm1, %zmm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,2,10,2,10,u,u,u,u,3,11,3,11,u,u,u,u,4,12,4,12,u,u,u,u,5,13,5,13,u,u,u,u,6,14,6,14,u,u,u,u,7,15,7,15]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm6, %zmm5, %zmm5
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm6, %zmm2, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm2 = zmm2[u,u,u,u,0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u]
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm5, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT: vpermq %zmm4, %zmm1, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpmovsxdq {{.*#+}} zmm4 = [134219776,151062785,167905794,184748803,201591812,218434821,235277830,252120839]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm4, %zmm1, %zmm1
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7]
-; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm4, %zmm0, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,8,0,8,u,u,u,u,1,9,1,9,u,u,u,u,18,26,18,26,u,u,u,u,19,27,19,27,u,u,u,u,36,44,u,u,u,u,u,u,37,45,u,u,u,u,u,u,54,62,u,u,u,u,u,u,55,63,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2}
; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
@@ -8620,230 +8612,336 @@ define void @store_i8_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512BW-NEXT: vmovdqa (%r10), %xmm1
-; AVX512BW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-NEXT: vmovdqa 16(%r10), %xmm13
-; AVX512BW-NEXT: vmovdqa64 32(%r10), %xmm22
-; AVX512BW-NEXT: vmovdqa64 48(%r10), %xmm19
-; AVX512BW-NEXT: vmovdqa (%rax), %xmm0
-; AVX512BW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-NEXT: vmovdqa64 16(%rax), %xmm16
-; AVX512BW-NEXT: vmovdqa64 32(%rax), %xmm23
-; AVX512BW-NEXT: vmovdqa64 48(%rax), %xmm20
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512BW-NEXT: vmovdqa64 16(%r9), %xmm17
-; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm25
-; AVX512BW-NEXT: vmovdqa64 48(%r9), %xmm21
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512BW-NEXT: vmovdqa64 16(%r8), %xmm18
-; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm26
-; AVX512BW-NEXT: vmovdqa64 48(%r8), %xmm24
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3,16,17,20,21,20,21,21,23,16,17,22,21,22,21,23,23]
-; AVX512BW-NEXT: vpermw %zmm1, %zmm10, %zmm4
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3,16,17,20,20,20,20,22,21,16,17,20,22,20,22,22,23]
+; AVX512BW-NEXT: vmovdqa (%r10), %xmm0
+; AVX512BW-NEXT: vmovdqa 16(%r10), %xmm12
+; AVX512BW-NEXT: vmovdqa64 32(%r10), %xmm16
+; AVX512BW-NEXT: vmovdqa 48(%r10), %xmm15
+; AVX512BW-NEXT: vmovdqa (%rax), %xmm2
+; AVX512BW-NEXT: vmovdqa 16(%rax), %xmm13
+; AVX512BW-NEXT: vmovdqa64 32(%rax), %xmm17
+; AVX512BW-NEXT: vmovdqa64 48(%rax), %xmm18
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3]
+; AVX512BW-NEXT: vpermw %ymm1, %ymm3, %ymm1
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512BW-NEXT: vmovdqa (%r9), %xmm4
+; AVX512BW-NEXT: vmovdqa64 48(%r9), %xmm19
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm5
+; AVX512BW-NEXT: vmovdqa64 48(%r8), %xmm21
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm7[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm7[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti128 $1, %xmm8, %ymm6, %ymm6
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3]
+; AVX512BW-NEXT: vpermw %ymm7, %ymm6, %ymm7
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm14
; AVX512BW-NEXT: movl $-2004318072, %eax # imm = 0x88888888
; AVX512BW-NEXT: kmovd %eax, %k1
-; AVX512BW-NEXT: vpermw %zmm0, %zmm9, %zmm4 {%k1}
-; AVX512BW-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512BW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-NEXT: vmovdqa64 48(%rsi), %xmm28
-; AVX512BW-NEXT: vmovdqa (%rdi), %xmm6
-; AVX512BW-NEXT: vmovdqa64 48(%rdi), %xmm30
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm0[8],xmm6[9],xmm0[9],xmm6[10],xmm0[10],xmm6[11],xmm0[11],xmm6[12],xmm0[12],xmm6[13],xmm0[13],xmm6[14],xmm0[14],xmm6[15],xmm0[15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm14 {%k1}
+; AVX512BW-NEXT: vmovdqa (%rsi), %xmm7
+; AVX512BW-NEXT: vmovdqa64 48(%rsi), %xmm24
+; AVX512BW-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512BW-NEXT: vmovdqa64 48(%rdi), %xmm27
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm10 = xmm1[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm8 = xmm0[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm8, %ymm1, %ymm1
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm0, %ymm8, %ymm0
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm3
-; AVX512BW-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512BW-NEXT: vmovdqa 48(%rcx), %xmm1
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = 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]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm12, %zmm12, %zmm14
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm12 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7,20,20,18,19,22,21,22,21,20,22,18,19,22,23,22,23]
+; AVX512BW-NEXT: vinserti128 $1, %xmm1, %ymm10, %ymm1
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
+; AVX512BW-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512BW-NEXT: vmovdqa64 48(%rcx), %xmm28
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512BW-NEXT: vmovdqa64 48(%rdx), %xmm29
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = 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: vpshufhw {{.*#+}} xmm11 = xmm20[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm22, %ymm11, %ymm11
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm22 = ymm11[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7]
+; AVX512BW-NEXT: vpermw %ymm20, %ymm11, %ymm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm22, %zmm20, %zmm20
; AVX512BW-NEXT: movl $572662306, %eax # imm = 0x22222222
; AVX512BW-NEXT: kmovd %eax, %k2
-; AVX512BW-NEXT: vpermw %zmm14, %zmm12, %zmm3 {%k2}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm14, %zmm14, %zmm15
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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-NEXT: vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512BW-NEXT: vpermw %zmm14, %zmm10, %zmm14
-; AVX512BW-NEXT: vpermw %zmm15, %zmm9, %zmm14 {%k1}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm27 = xmm15[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm27[0],zero,zero,zero,xmm27[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm29 = xmm15[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm29[0],zero,zero,zero,xmm29[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm29, %ymm27, %ymm27
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
+; AVX512BW-NEXT: vmovdqu16 %zmm20, %zmm1 {%k2}
+; AVX512BW-NEXT: movw $-21846, %ax # imm = 0xAAAA
+; AVX512BW-NEXT: kmovd %eax, %k3
+; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm1 {%k3}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3],xmm18[4],xmm15[4],xmm18[5],xmm15[5],xmm18[6],xmm15[6],xmm18[7],xmm15[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm14[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm14[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm22, %ymm20, %ymm20
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm14, %ymm3, %ymm14
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm14, %zmm14
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm21[0],xmm19[0],xmm21[1],xmm19[1],xmm21[2],xmm19[2],xmm21[3],xmm19[3],xmm21[4],xmm19[4],xmm21[5],xmm19[5],xmm21[6],xmm19[6],xmm21[7],xmm19[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm20[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm22, %ymm22
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm22 = ymm22[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm20, %ymm6, %ymm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm22, %zmm20, %zmm23
+; AVX512BW-NEXT: vmovdqu16 %zmm14, %zmm23 {%k1}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm27[0],xmm24[0],xmm27[1],xmm24[1],xmm27[2],xmm24[2],xmm27[3],xmm24[3],xmm27[4],xmm24[4],xmm27[5],xmm24[5],xmm27[6],xmm24[6],xmm27[7],xmm24[7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm14[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm22 = xmm14[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm22[0],zero,zero,zero,xmm22[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm22, %ymm20, %ymm20
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm14, %ymm22, %ymm14
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm14, %zmm14
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm20[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm25, %ymm22, %ymm25
+; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm20, %ymm11, %ymm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm20
+; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm25
+; AVX512BW-NEXT: vmovdqu16 %zmm20, %zmm14 {%k2}
+; AVX512BW-NEXT: vmovdqa64 32(%rsi), %xmm20
+; AVX512BW-NEXT: vmovdqa32 %zmm23, %zmm14 {%k3}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm18[8],xmm15[8],xmm18[9],xmm15[9],xmm18[10],xmm15[10],xmm18[11],xmm15[11],xmm18[12],xmm15[12],xmm18[13],xmm15[13],xmm18[14],xmm15[14],xmm18[15],xmm15[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm18 = xmm15[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm15[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm18, %ymm18
+; AVX512BW-NEXT: vmovdqa64 32(%rdi), %xmm23
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm18 = ymm18[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm15, %ymm3, %ymm15
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm18, %zmm15, %zmm15
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm21[8],xmm19[8],xmm21[9],xmm19[9],xmm21[10],xmm19[10],xmm21[11],xmm19[11],xmm21[12],xmm19[12],xmm21[13],xmm19[13],xmm21[14],xmm19[14],xmm21[15],xmm19[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm19 = xmm18[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm18[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm21, %ymm19, %ymm21
+; AVX512BW-NEXT: vmovdqa64 32(%rcx), %xmm19
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm21[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm18, %ymm6, %ymm18
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512BW-NEXT: vmovdqa64 32(%rdx), %xmm26
+; AVX512BW-NEXT: vmovdqu16 %zmm15, %zmm18 {%k1}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm27[8],xmm24[8],xmm27[9],xmm24[9],xmm27[10],xmm24[10],xmm27[11],xmm24[11],xmm27[12],xmm24[12],xmm27[13],xmm24[13],xmm27[14],xmm24[14],xmm27[15],xmm24[15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm15[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm15[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
; AVX512BW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,1,1]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm15 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm15, %ymm29, %ymm15
-; AVX512BW-NEXT: vmovdqa 48(%rdx), %xmm0
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm27, %zmm15, %zmm15
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm27 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm27, %zmm27, %zmm27
-; AVX512BW-NEXT: vpermw %zmm27, %zmm12, %zmm15 {%k2}
-; AVX512BW-NEXT: vmovdqa64 32(%rsi), %xmm29
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15]
-; AVX512BW-NEXT: vmovdqa64 32(%rdi), %xmm31
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = 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-NEXT: vmovdqa64 32(%rcx), %xmm27
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm19, %zmm19, %zmm21
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm20, %zmm20, %zmm19
-; AVX512BW-NEXT: vpermw %zmm19, %zmm10, %zmm19
-; AVX512BW-NEXT: vpermw %zmm21, %zmm9, %zmm19 {%k1}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm20[2,3,2,3]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm15, %ymm24, %ymm15
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm15, %zmm15
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm21[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm27, %ymm24, %ymm24
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm24 = ymm24[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm21, %ymm11, %ymm21
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm24, %zmm21, %zmm21
+; AVX512BW-NEXT: vmovdqu16 %zmm21, %zmm15 {%k2}
+; AVX512BW-NEXT: vmovdqa32 %zmm18, %zmm15 {%k3}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm17[0],xmm16[0],xmm17[1],xmm16[1],xmm17[2],xmm16[2],xmm17[3],xmm16[3],xmm17[4],xmm16[4],xmm17[5],xmm16[5],xmm17[6],xmm16[6],xmm17[7],xmm16[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm18[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm18[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm21[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm18, %ymm3, %ymm18
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm25[0],xmm22[0],xmm25[1],xmm22[1],xmm25[2],xmm22[2],xmm25[3],xmm22[3],xmm25[4],xmm22[4],xmm25[5],xmm22[5],xmm25[6],xmm22[6],xmm25[7],xmm22[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm21[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm27, %ymm24, %ymm24
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm24 = ymm24[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm21, %ymm6, %ymm21
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm24, %zmm21, %zmm27
+; AVX512BW-NEXT: vmovdqu16 %zmm18, %zmm27 {%k1}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm23[0],xmm20[0],xmm23[1],xmm20[1],xmm23[2],xmm20[2],xmm23[3],xmm20[3],xmm23[4],xmm20[4],xmm23[5],xmm20[5],xmm23[6],xmm20[6],xmm23[7],xmm20[7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm18[2,3,2,3]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm20[3,3,3,3]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm18[3,3,3,3]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
; AVX512BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm20[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm18 = xmm18[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm18, %ymm24, %ymm18
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm26[0],xmm19[0],xmm26[1],xmm19[1],xmm26[2],xmm19[2],xmm26[3],xmm19[3],xmm26[4],xmm19[4],xmm26[5],xmm19[5],xmm26[6],xmm19[6],xmm26[7],xmm19[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm28 = xmm21[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm28, %ymm24, %ymm28
+; AVX512BW-NEXT: vmovdqa64 16(%r9), %xmm24
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm28 = ymm28[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm21, %ymm11, %ymm21
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm28, %zmm21, %zmm21
+; AVX512BW-NEXT: vmovdqa64 16(%r8), %xmm28
+; AVX512BW-NEXT: vmovdqu16 %zmm21, %zmm18 {%k2}
+; AVX512BW-NEXT: vmovdqa64 16(%rsi), %xmm21
+; AVX512BW-NEXT: vmovdqa32 %zmm27, %zmm18 {%k3}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm17 = xmm16[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm16[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm27, %ymm17, %ymm17
+; AVX512BW-NEXT: vmovdqa64 16(%rdi), %xmm27
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm16, %ymm3, %ymm16
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm17, %zmm16, %zmm16
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm25[8],xmm22[8],xmm25[9],xmm22[9],xmm25[10],xmm22[10],xmm25[11],xmm22[11],xmm25[12],xmm22[12],xmm25[13],xmm22[13],xmm25[14],xmm22[14],xmm25[15],xmm22[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm17 = xmm22[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm22[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm25, %ymm17, %ymm25
+; AVX512BW-NEXT: vmovdqa64 16(%rcx), %xmm17
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm22, %ymm6, %ymm22
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm25, %zmm22, %zmm25
+; AVX512BW-NEXT: vmovdqa64 16(%rdx), %xmm22
+; AVX512BW-NEXT: vmovdqu16 %zmm16, %zmm25 {%k1}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm23[8],xmm20[8],xmm23[9],xmm20[9],xmm23[10],xmm20[10],xmm23[11],xmm20[11],xmm23[12],xmm20[12],xmm23[13],xmm20[13],xmm23[14],xmm20[14],xmm23[15],xmm20[15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm16[2,3,2,3]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm20, %ymm24, %ymm20
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm21, %zmm20, %zmm20
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpermw %zmm0, %zmm12, %zmm20 {%k2}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512BW-NEXT: vpermw %zmm1, %zmm10, %zmm21
-; AVX512BW-NEXT: vpermw %zmm0, %zmm9, %zmm21 {%k1}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm0[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm24, %ymm1, %ymm1
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm24, %ymm0
-; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm24
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm27[0],xmm2[1],xmm27[1],xmm2[2],xmm27[2],xmm2[3],xmm27[3],xmm2[4],xmm27[4],xmm2[5],xmm27[5],xmm2[6],xmm27[6],xmm2[7],xmm27[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpermw %zmm0, %zmm12, %zmm24 {%k2}
-; AVX512BW-NEXT: vmovdqa64 16(%rsi), %xmm28
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
-; AVX512BW-NEXT: vmovdqa64 16(%rdi), %xmm30
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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: vmovdqa64 16(%rcx), %xmm25
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512BW-NEXT: vpermw %zmm1, %zmm10, %zmm22
-; AVX512BW-NEXT: vpermw %zmm0, %zmm9, %zmm22 {%k1}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm31[8],xmm29[8],xmm31[9],xmm29[9],xmm31[10],xmm29[10],xmm31[11],xmm29[11],xmm31[12],xmm29[12],xmm31[13],xmm29[13],xmm31[14],xmm29[14],xmm31[15],xmm29[15]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm0[3,3,3,3]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm16[3,3,3,3]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm1, %ymm1
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm23, %ymm0
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm23
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm27[8],xmm2[9],xmm27[9],xmm2[10],xmm27[10],xmm2[11],xmm27[11],xmm2[12],xmm27[12],xmm2[13],xmm27[13],xmm2[14],xmm27[14],xmm2[15],xmm27[15]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpermw %zmm0, %zmm12, %zmm23 {%k2}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm16[0],xmm13[0],xmm16[1],xmm13[1],xmm16[2],xmm13[2],xmm16[3],xmm13[3],xmm16[4],xmm13[4],xmm16[5],xmm13[5],xmm16[6],xmm13[6],xmm16[7],xmm13[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512BW-NEXT: vpermw %zmm1, %zmm10, %zmm26
-; AVX512BW-NEXT: vpermw %zmm0, %zmm9, %zmm26 {%k1}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm16 = xmm16[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm16 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm16, %ymm23, %ymm16
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm16, %zmm16
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm26[8],xmm19[8],xmm26[9],xmm19[9],xmm26[10],xmm19[10],xmm26[11],xmm19[11],xmm26[12],xmm19[12],xmm26[13],xmm19[13],xmm26[14],xmm19[14],xmm26[15],xmm19[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm19[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm19[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm19, %ymm11, %ymm19
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm19
+; AVX512BW-NEXT: vmovdqu16 %zmm19, %zmm16 {%k2}
+; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm16 {%k3}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm19[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm19[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm19, %ymm3, %ymm19
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm19
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm28[0],xmm24[0],xmm28[1],xmm24[1],xmm28[2],xmm24[2],xmm28[3],xmm24[3],xmm28[4],xmm24[4],xmm28[5],xmm24[5],xmm28[6],xmm24[6],xmm28[7],xmm24[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm20[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm20[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm25, %ymm23, %ymm23
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm23 = ymm23[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm20, %ymm6, %ymm20
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512BW-NEXT: vmovdqu16 %zmm19, %zmm20 {%k1}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm27[0],xmm21[0],xmm27[1],xmm21[1],xmm27[2],xmm21[2],xmm27[3],xmm21[3],xmm27[4],xmm21[4],xmm27[5],xmm21[5],xmm27[6],xmm21[6],xmm27[7],xmm21[7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm19[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm25 = xmm19[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm25[0],zero,zero,zero,xmm25[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm25, %ymm23, %ymm23
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm19[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm19 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm19, %ymm25, %ymm19
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm23, %zmm19, %zmm19
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm22[0],xmm17[0],xmm22[1],xmm17[1],xmm22[2],xmm17[2],xmm22[3],xmm17[3],xmm22[4],xmm17[4],xmm22[5],xmm17[5],xmm22[6],xmm17[6],xmm22[7],xmm17[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm23[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm26 = xmm23[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm26, %ymm25, %ymm25
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm23, %ymm11, %ymm23
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm25, %zmm23, %zmm23
+; AVX512BW-NEXT: vmovdqu16 %zmm23, %zmm19 {%k2}
+; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm19 {%k3}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm12[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm12[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm20, %ymm13, %ymm13
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm12, %ymm3, %ymm12
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm28[8],xmm24[8],xmm28[9],xmm24[9],xmm28[10],xmm24[10],xmm28[11],xmm24[11],xmm28[12],xmm24[12],xmm28[13],xmm24[13],xmm28[14],xmm24[14],xmm28[15],xmm24[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm13[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm13[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vpermw %ymm13, %ymm6, %ymm13
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm13, %zmm13
+; AVX512BW-NEXT: vmovdqu16 %zmm12, %zmm13 {%k1}
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm27[8],xmm21[8],xmm27[9],xmm21[9],xmm27[10],xmm21[10],xmm27[11],xmm21[11],xmm27[12],xmm21[12],xmm27[13],xmm21[13],xmm27[14],xmm21[14],xmm27[15],xmm21[15]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm12[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm12[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm21, %ymm20, %ymm20
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,1,1]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm12, %ymm21, %ymm12
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm12, %zmm12
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm22[8],xmm17[8],xmm22[9],xmm17[9],xmm22[10],xmm17[10],xmm22[11],xmm17[11],xmm22[12],xmm17[12],xmm22[13],xmm17[13],xmm22[14],xmm17[14],xmm22[15],xmm17[15]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm17[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm17[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti32x4 $1, %xmm21, %ymm20, %ymm20
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vpermw %ymm17, %ymm11, %ymm17
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm20, %zmm17, %zmm17
+; AVX512BW-NEXT: vmovdqu16 %zmm17, %zmm12 {%k2}
+; AVX512BW-NEXT: vmovdqa32 %zmm13, %zmm12 {%k3}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm0[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti128 $1, %xmm13, %ymm2, %ymm2
+; AVX512BW-NEXT: vpermw %ymm0, %ymm3, %ymm0
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512BW-NEXT: vpermw %ymm2, %ymm6, %ymm2
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm2 {%k1}
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
+; AVX512BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
+; AVX512BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512BW-NEXT: vmovdqa 16(%rdx), %xmm2
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm27
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm25[0],xmm2[1],xmm25[1],xmm2[2],xmm25[2],xmm2[3],xmm25[3],xmm2[4],xmm25[4],xmm2[5],xmm25[5],xmm2[6],xmm25[6],xmm2[7],xmm25[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512BW-NEXT: vpermw %zmm0, %zmm12, %zmm27 {%k2}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm16[8],xmm13[8],xmm16[9],xmm13[9],xmm16[10],xmm13[10],xmm16[11],xmm13[11],xmm16[12],xmm13[12],xmm16[13],xmm13[13],xmm16[14],xmm13[14],xmm16[15],xmm13[15]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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: vinserti32x4 $2, %xmm0, %zmm0, %zmm13
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm0
-; AVX512BW-NEXT: vpermw %zmm0, %zmm10, %zmm0
-; AVX512BW-NEXT: vpermw %zmm13, %zmm9, %zmm0 {%k1}
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
-; AVX512BW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512BW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
-; AVX512BW-NEXT: # xmm13 = xmm13[0],mem[0],xmm13[1],mem[1],xmm13[2],mem[2],xmm13[3],mem[3],xmm13[4],mem[4],xmm13[5],mem[5],xmm13[6],mem[6],xmm13[7],mem[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm5, %zmm5, %zmm5
-; AVX512BW-NEXT: vpermw %zmm5, %zmm10, %zmm5
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm7
-; AVX512BW-NEXT: vpermw %zmm7, %zmm9, %zmm5 {%k1}
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm1[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm9, %ymm7, %ymm7
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm1, %ymm9, %ymm1
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm25[8],xmm2[9],xmm25[9],xmm2[10],xmm25[10],xmm2[11],xmm25[11],xmm2[12],xmm25[12],xmm2[13],xmm25[13],xmm2[14],xmm25[14],xmm2[15],xmm25[15]
-; AVX512BW-NEXT: vinserti32x4 $2, %xmm2, %zmm2, %zmm2
-; AVX512BW-NEXT: vpermw %zmm2, %zmm12, %zmm1 {%k2}
-; AVX512BW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm2 # 16-byte Folded Reload
-; AVX512BW-NEXT: # xmm2 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3],xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[2,3,2,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm2[3,3,3,3]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
-; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm7, %ymm2
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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-NEXT: vinserti32x4 $2, %xmm6, %zmm6, %zmm6
-; AVX512BW-NEXT: vpermw %zmm6, %zmm12, %zmm2 {%k2}
-; AVX512BW-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512BW-NEXT: kmovd %eax, %k1
-; AVX512BW-NEXT: vmovdqa32 %zmm4, %zmm3 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm15 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm20 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm24 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm22, %zmm23 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm26, %zmm27 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512BW-NEXT: vmovdqa32 %zmm5, %zmm2 {%k1}
+; AVX512BW-NEXT: vinserti128 $1, %xmm0, %ymm4, %ymm0
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
+; AVX512BW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
+; AVX512BW-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4
+; AVX512BW-NEXT: vpermw %ymm3, %ymm11, %ymm3
+; AVX512BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7]
+; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm0 {%k2}
+; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm27, 128(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm23, 320(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm24, 256(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm20, 448(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm15, 384(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm12, 192(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm19, 128(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm16, 320(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm18, 256(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm15, 448(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm1, 64(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
@@ -9024,230 +9122,336 @@ define void @store_i8_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW: # %bb.0:
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-BW-NEXT: vmovdqa 16(%r10), %xmm13
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r10), %xmm22
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%r10), %xmm19
-; AVX512DQ-BW-NEXT: vmovdqa (%rax), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%rax), %xmm16
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rax), %xmm23
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%rax), %xmm20
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%r9), %xmm17
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm25
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%r9), %xmm21
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%r8), %xmm18
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm26
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%r8), %xmm24
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3,16,17,20,21,20,21,21,23,16,17,22,21,22,21,23,23]
-; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm10, %zmm4
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3,16,17,20,20,20,20,22,21,16,17,20,22,20,22,22,23]
+; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm0
+; AVX512DQ-BW-NEXT: vmovdqa 16(%r10), %xmm12
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r10), %xmm16
+; AVX512DQ-BW-NEXT: vmovdqa 48(%r10), %xmm15
+; AVX512DQ-BW-NEXT: vmovdqa (%rax), %xmm2
+; AVX512DQ-BW-NEXT: vmovdqa 16(%rax), %xmm13
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rax), %xmm17
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%rax), %xmm18
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,0,0,0,4,5,2,1,0,2,0,2,4,5,2,3]
+; AVX512DQ-BW-NEXT: vpermw %ymm1, %ymm3, %ymm1
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm4
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%r9), %xmm19
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%r8), %xmm21
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm7[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm7[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm8, %ymm6, %ymm6
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,1,0,1,4,5,1,3,2,1,2,1,4,5,3,3]
+; AVX512DQ-BW-NEXT: vpermw %ymm7, %ymm6, %ymm7
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm14
; AVX512DQ-BW-NEXT: movl $-2004318072, %eax # imm = 0x88888888
; AVX512DQ-BW-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm9, %zmm4 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%rsi), %xmm28
-; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm6
-; AVX512DQ-BW-NEXT: vmovdqa64 48(%rdi), %xmm30
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm0[8],xmm6[9],xmm0[9],xmm6[10],xmm0[10],xmm6[11],xmm0[11],xmm6[12],xmm0[12],xmm6[13],xmm0[13],xmm6[14],xmm0[14],xmm6[15],xmm0[15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm14 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm7
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%rsi), %xmm24
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%rdi), %xmm27
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm10 = xmm1[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm10, %ymm9, %ymm9
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm8 = xmm0[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm8, %ymm1, %ymm1
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm0, %ymm8, %ymm0
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm3
-; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512DQ-BW-NEXT: vmovdqa 48(%rcx), %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = 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-BW-NEXT: vinserti32x4 $2, %xmm12, %zmm12, %zmm14
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm12 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7,20,20,18,19,22,21,22,21,20,22,18,19,22,23,22,23]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm1, %ymm10, %ymm1
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
+; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%rcx), %xmm28
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512DQ-BW-NEXT: vmovdqa64 48(%rdx), %xmm29
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = 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: vpshufhw {{.*#+}} xmm11 = xmm20[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm22, %ymm11, %ymm11
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm22 = ymm11[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,0,2,1,2,1,6,7,0,2,2,3,2,3,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm20, %ymm11, %ymm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm22, %zmm20, %zmm20
; AVX512DQ-BW-NEXT: movl $572662306, %eax # imm = 0x22222222
; AVX512DQ-BW-NEXT: kmovd %eax, %k2
-; AVX512DQ-BW-NEXT: vpermw %zmm14, %zmm12, %zmm3 {%k2}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm14, %zmm14, %zmm15
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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-NEXT: vinserti32x4 $2, %xmm14, %zmm14, %zmm14
-; AVX512DQ-BW-NEXT: vpermw %zmm14, %zmm10, %zmm14
-; AVX512DQ-BW-NEXT: vpermw %zmm15, %zmm9, %zmm14 {%k1}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm27 = xmm15[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm27[0],zero,zero,zero,xmm27[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm29 = xmm15[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm29[0],zero,zero,zero,xmm29[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm29, %ymm27, %ymm27
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm29 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm20, %zmm1 {%k2}
+; AVX512DQ-BW-NEXT: movw $-21846, %ax # imm = 0xAAAA
+; AVX512DQ-BW-NEXT: kmovd %eax, %k3
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm14, %zmm1 {%k3}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3],xmm18[4],xmm15[4],xmm18[5],xmm15[5],xmm18[6],xmm15[6],xmm18[7],xmm15[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm14[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm14[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm22, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm14, %ymm3, %ymm14
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm14, %zmm14
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm21[0],xmm19[0],xmm21[1],xmm19[1],xmm21[2],xmm19[2],xmm21[3],xmm19[3],xmm21[4],xmm19[4],xmm21[5],xmm19[5],xmm21[6],xmm19[6],xmm21[7],xmm19[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm20[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm22, %ymm22
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm22 = ymm22[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm20, %ymm6, %ymm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm22, %zmm20, %zmm23
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm14, %zmm23 {%k1}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm27[0],xmm24[0],xmm27[1],xmm24[1],xmm27[2],xmm24[2],xmm27[3],xmm24[3],xmm27[4],xmm24[4],xmm27[5],xmm24[5],xmm27[6],xmm24[6],xmm27[7],xmm24[7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm14[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm22 = xmm14[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm22[0],zero,zero,zero,xmm22[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm22, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm22 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm14, %ymm22, %ymm14
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm14, %zmm14
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm22 = xmm20[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm20[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm25, %ymm22, %ymm25
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm20, %ymm11, %ymm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm25
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm20, %zmm14 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %xmm20
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm23, %zmm14 {%k3}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm18[8],xmm15[8],xmm18[9],xmm15[9],xmm18[10],xmm15[10],xmm18[11],xmm15[11],xmm18[12],xmm15[12],xmm18[13],xmm15[13],xmm18[14],xmm15[14],xmm18[15],xmm15[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm18 = xmm15[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm15[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm18, %ymm18
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %xmm23
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm18 = ymm18[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm15, %ymm3, %ymm15
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm18, %zmm15, %zmm15
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm21[8],xmm19[8],xmm21[9],xmm19[9],xmm21[10],xmm19[10],xmm21[11],xmm19[11],xmm21[12],xmm19[12],xmm21[13],xmm19[13],xmm21[14],xmm19[14],xmm21[15],xmm19[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm19 = xmm18[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm18[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm21, %ymm19, %ymm21
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %xmm19
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm21[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm18, %ymm6, %ymm18
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %xmm26
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm15, %zmm18 {%k1}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm27[8],xmm24[8],xmm27[9],xmm24[9],xmm27[10],xmm24[10],xmm27[11],xmm24[11],xmm27[12],xmm24[12],xmm27[13],xmm24[13],xmm27[14],xmm24[14],xmm27[15],xmm24[15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm15[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm15[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,1,1]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm15 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm15, %ymm29, %ymm15
-; AVX512DQ-BW-NEXT: vmovdqa 48(%rdx), %xmm0
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm27, %zmm15, %zmm15
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm27 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm27, %zmm27, %zmm27
-; AVX512DQ-BW-NEXT: vpermw %zmm27, %zmm12, %zmm15 {%k2}
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %xmm29
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15]
-; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %xmm31
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = 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-NEXT: vmovdqa64 32(%rcx), %xmm27
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm19, %zmm19, %zmm21
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm20, %zmm20, %zmm19
-; AVX512DQ-BW-NEXT: vpermw %zmm19, %zmm10, %zmm19
-; AVX512DQ-BW-NEXT: vpermw %zmm21, %zmm9, %zmm19 {%k1}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm20[2,3,2,3]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm15, %ymm24, %ymm15
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm15, %zmm15
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm21[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm27, %ymm24, %ymm24
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm24 = ymm24[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm21, %ymm11, %ymm21
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm24, %zmm21, %zmm21
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm21, %zmm15 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm18, %zmm15 {%k3}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm17[0],xmm16[0],xmm17[1],xmm16[1],xmm17[2],xmm16[2],xmm17[3],xmm16[3],xmm17[4],xmm16[4],xmm17[5],xmm16[5],xmm17[6],xmm16[6],xmm17[7],xmm16[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm18[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm18[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm21 = ymm21[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm18, %ymm3, %ymm18
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm25[0],xmm22[0],xmm25[1],xmm22[1],xmm25[2],xmm22[2],xmm25[3],xmm22[3],xmm25[4],xmm22[4],xmm25[5],xmm22[5],xmm25[6],xmm22[6],xmm25[7],xmm22[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm21[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm27, %ymm24, %ymm24
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm24 = ymm24[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm21, %ymm6, %ymm21
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm24, %zmm21, %zmm27
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm18, %zmm27 {%k1}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm23[0],xmm20[0],xmm23[1],xmm20[1],xmm23[2],xmm20[2],xmm23[3],xmm20[3],xmm23[4],xmm20[4],xmm23[5],xmm20[5],xmm23[6],xmm20[6],xmm23[7],xmm20[7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm18[2,3,2,3]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm20[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm18[3,3,3,3]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm24, %ymm21, %ymm21
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm20[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm18 = xmm18[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm18, %ymm24, %ymm18
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm18, %zmm18
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm26[0],xmm19[0],xmm26[1],xmm19[1],xmm26[2],xmm19[2],xmm26[3],xmm19[3],xmm26[4],xmm19[4],xmm26[5],xmm19[5],xmm26[6],xmm19[6],xmm26[7],xmm19[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm24 = xmm21[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm28 = xmm21[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm28, %ymm24, %ymm28
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%r9), %xmm24
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm28 = ymm28[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm21, %ymm11, %ymm21
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm28, %zmm21, %zmm21
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%r8), %xmm28
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm21, %zmm18 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%rsi), %xmm21
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm27, %zmm18 {%k3}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm17 = xmm16[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm27 = xmm16[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm27, %ymm17, %ymm17
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdi), %xmm27
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm16, %ymm3, %ymm16
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm17, %zmm16, %zmm16
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm25[8],xmm22[8],xmm25[9],xmm22[9],xmm25[10],xmm22[10],xmm25[11],xmm22[11],xmm25[12],xmm22[12],xmm25[13],xmm22[13],xmm25[14],xmm22[14],xmm25[15],xmm22[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm17 = xmm22[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm22[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm25, %ymm17, %ymm25
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%rcx), %xmm17
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm22, %ymm6, %ymm22
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm25, %zmm22, %zmm25
+; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdx), %xmm22
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm16, %zmm25 {%k1}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm23[8],xmm20[8],xmm23[9],xmm20[9],xmm23[10],xmm20[10],xmm23[11],xmm20[11],xmm23[12],xmm20[12],xmm23[13],xmm20[13],xmm23[14],xmm20[14],xmm23[15],xmm20[15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm16[2,3,2,3]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm20, %ymm24, %ymm20
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm21, %zmm20, %zmm20
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm12, %zmm20 {%k2}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm10, %zmm21
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm9, %zmm21 {%k1}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm24 = xmm0[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm24[0],zero,zero,zero,xmm24[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm24, %ymm1, %ymm1
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm0, %ymm24, %ymm0
-; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm24
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm27[0],xmm2[1],xmm27[1],xmm2[2],xmm27[2],xmm2[3],xmm27[3],xmm2[4],xmm27[4],xmm2[5],xmm27[5],xmm2[6],xmm27[6],xmm2[7],xmm27[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm12, %zmm24 {%k2}
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%rsi), %xmm28
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15]
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdi), %xmm30
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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: vmovdqa64 16(%rcx), %xmm25
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm10, %zmm22
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm9, %zmm22 {%k1}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm31[8],xmm29[8],xmm31[9],xmm29[9],xmm31[10],xmm29[10],xmm31[11],xmm29[11],xmm31[12],xmm29[12],xmm31[13],xmm29[13],xmm31[14],xmm29[14],xmm31[15],xmm29[15]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm0[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm16[3,3,3,3]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm1, %ymm1
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm0, %ymm23, %ymm0
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm23
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm27[8],xmm2[9],xmm27[9],xmm2[10],xmm27[10],xmm2[11],xmm27[11],xmm2[12],xmm27[12],xmm2[13],xmm27[13],xmm2[14],xmm27[14],xmm2[15],xmm27[15]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm12, %zmm23 {%k2}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm16[0],xmm13[0],xmm16[1],xmm13[1],xmm16[2],xmm13[2],xmm16[3],xmm13[3],xmm16[4],xmm13[4],xmm16[5],xmm13[5],xmm16[6],xmm13[6],xmm16[7],xmm13[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm10, %zmm26
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm9, %zmm26 {%k1}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm16 = xmm16[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm16 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm16, %ymm23, %ymm16
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm16, %zmm16
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm26[8],xmm19[8],xmm26[9],xmm19[9],xmm26[10],xmm19[10],xmm26[11],xmm19[11],xmm26[12],xmm19[12],xmm26[13],xmm19[13],xmm26[14],xmm19[14],xmm26[15],xmm19[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm19[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm19[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm19, %ymm11, %ymm19
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm19
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm19, %zmm16 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm25, %zmm16 {%k3}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm19[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm19[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm19, %ymm3, %ymm19
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm19
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm28[0],xmm24[0],xmm28[1],xmm24[1],xmm28[2],xmm24[2],xmm28[3],xmm24[3],xmm28[4],xmm24[4],xmm28[5],xmm24[5],xmm28[6],xmm24[6],xmm28[7],xmm24[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm20[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm20[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm25, %ymm23, %ymm23
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm23 = ymm23[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm20, %ymm6, %ymm20
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm20
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm19, %zmm20 {%k1}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm27[0],xmm21[0],xmm27[1],xmm21[1],xmm27[2],xmm21[2],xmm27[3],xmm21[3],xmm27[4],xmm21[4],xmm27[5],xmm21[5],xmm27[6],xmm21[6],xmm27[7],xmm21[7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm23 = xmm19[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm23 = xmm23[0],zero,zero,zero,xmm23[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm25 = xmm19[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm25[0],zero,zero,zero,xmm25[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm25, %ymm23, %ymm23
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm19[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm19 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm19, %ymm25, %ymm19
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm23, %zmm19, %zmm19
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm22[0],xmm17[0],xmm22[1],xmm17[1],xmm22[2],xmm17[2],xmm22[3],xmm17[3],xmm22[4],xmm17[4],xmm22[5],xmm17[5],xmm22[6],xmm17[6],xmm22[7],xmm17[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm25 = xmm23[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm26 = xmm23[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm26, %ymm25, %ymm25
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm23, %ymm11, %ymm23
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm25, %zmm23, %zmm23
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm23, %zmm19 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm20, %zmm19 {%k3}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm12[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm12[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm20, %ymm13, %ymm13
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm12, %ymm3, %ymm12
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm28[8],xmm24[8],xmm28[9],xmm24[9],xmm28[10],xmm24[10],xmm28[11],xmm24[11],xmm28[12],xmm24[12],xmm28[13],xmm24[13],xmm28[14],xmm24[14],xmm28[15],xmm24[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm13[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm23 = xmm13[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm23, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm13, %ymm6, %ymm13
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm13, %zmm13
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm12, %zmm13 {%k1}
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm27[8],xmm21[8],xmm27[9],xmm21[9],xmm27[10],xmm21[10],xmm27[11],xmm21[11],xmm27[12],xmm21[12],xmm27[13],xmm21[13],xmm27[14],xmm21[14],xmm27[15],xmm21[15]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm20 = xmm12[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm20 = xmm20[0],zero,zero,zero,xmm20[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm21 = xmm12[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm21, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,1,1]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm12, %ymm21, %ymm12
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm12, %zmm12
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm22[8],xmm17[8],xmm22[9],xmm17[9],xmm22[10],xmm17[10],xmm22[11],xmm17[11],xmm22[12],xmm17[12],xmm22[13],xmm17[13],xmm22[14],xmm17[14],xmm22[15],xmm17[15]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm20 = xmm17[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm21 = xmm17[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti32x4 $1, %xmm21, %ymm20, %ymm20
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm20 = ymm20[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vpermw %ymm17, %ymm11, %ymm17
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm20, %zmm17, %zmm17
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm17, %zmm12 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm13, %zmm12 {%k3}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm0[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm13, %ymm2, %ymm2
+; AVX512DQ-BW-NEXT: vpermw %ymm0, %ymm3, %ymm0
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,7]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vpermw %ymm2, %ymm6, %ymm2
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm0, %zmm2 {%k1}
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX512DQ-BW-NEXT: vmovdqa 16(%rdx), %xmm2
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm27
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm25[0],xmm2[1],xmm25[1],xmm2[2],xmm25[2],xmm2[3],xmm25[3],xmm2[4],xmm25[4],xmm2[5],xmm25[5],xmm2[6],xmm25[6],xmm2[7],xmm25[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm0, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm12, %zmm27 {%k2}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm16[8],xmm13[8],xmm16[9],xmm13[9],xmm16[10],xmm13[10],xmm16[11],xmm13[11],xmm16[12],xmm13[12],xmm16[13],xmm13[13],xmm16[14],xmm13[14],xmm16[15],xmm13[15]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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: vinserti32x4 $2, %xmm0, %zmm0, %zmm13
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm1, %zmm1, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm10, %zmm0
-; AVX512DQ-BW-NEXT: vpermw %zmm13, %zmm9, %zmm0 {%k1}
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15]
-; AVX512DQ-BW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512DQ-BW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
-; AVX512DQ-BW-NEXT: # xmm13 = xmm13[0],mem[0],xmm13[1],mem[1],xmm13[2],mem[2],xmm13[3],mem[3],xmm13[4],mem[4],xmm13[5],mem[5],xmm13[6],mem[6],xmm13[7],mem[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm5, %zmm5, %zmm5
-; AVX512DQ-BW-NEXT: vpermw %zmm5, %zmm10, %zmm5
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm13, %zmm13, %zmm7
-; AVX512DQ-BW-NEXT: vpermw %zmm7, %zmm9, %zmm5 {%k1}
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm1[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm9, %ymm7, %ymm7
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm1, %ymm9, %ymm1
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm25[8],xmm2[9],xmm25[9],xmm2[10],xmm25[10],xmm2[11],xmm25[11],xmm2[12],xmm25[12],xmm2[13],xmm25[13],xmm2[14],xmm25[14],xmm2[15],xmm25[15]
-; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm2, %zmm2, %zmm2
-; AVX512DQ-BW-NEXT: vpermw %zmm2, %zmm12, %zmm1 {%k2}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm2 # 16-byte Folded Reload
-; AVX512DQ-BW-NEXT: # xmm2 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3],xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[2,3,2,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm2[3,3,3,3]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
-; AVX512DQ-BW-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero
-; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm7, %ymm2
-; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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-NEXT: vinserti32x4 $2, %xmm6, %zmm6, %zmm6
-; AVX512DQ-BW-NEXT: vpermw %zmm6, %zmm12, %zmm2 {%k2}
-; AVX512DQ-BW-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512DQ-BW-NEXT: kmovd %eax, %k1
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm4, %zmm3 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm14, %zmm15 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm19, %zmm20 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm24 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm22, %zmm23 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm26, %zmm27 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm5, %zmm2 {%k1}
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm0, %ymm4, %ymm0
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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-BW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5]
+; AVX512DQ-BW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7]
+; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4
+; AVX512DQ-BW-NEXT: vpermw %ymm3, %ymm11, %ymm3
+; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7]
+; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512DQ-BW-NEXT: vmovdqu16 %zmm3, %zmm0 {%k2}
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k3}
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, 128(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, 320(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, 256(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 448(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 384(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 192(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 128(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 320(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 256(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 448(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 64(%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/widen_bitcnt.ll b/llvm/test/CodeGen/X86/widen_bitcnt.ll
index 18cb5e1b86ec3..541dfb54e96d2 100644
--- a/llvm/test/CodeGen/X86/widen_bitcnt.ll
+++ b/llvm/test/CodeGen/X86/widen_bitcnt.ll
@@ -629,71 +629,45 @@ define <8 x i32> @widen_ctlz_v2i32_v8i32(<2 x i32> %a0, <2 x i32> %a1, <2 x i32>
; AVX2: # %bb.0:
; AVX2-NEXT: # kill: def $xmm1 killed $xmm1 def $ymm1
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX2-NEXT: vmovq {{.*#+}} xmm5 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
-; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm4
-; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm6
-; AVX2-NEXT: vpbroadcastb {{.*#+}} xmm7 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT: vpand %xmm7, %xmm6, %xmm8
-; AVX2-NEXT: vpxor %xmm6, %xmm6, %xmm6
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm8, %xmm9
-; AVX2-NEXT: vpand %xmm4, %xmm9, %xmm4
-; AVX2-NEXT: vpshufb %xmm8, %xmm5, %xmm8
-; AVX2-NEXT: vpaddb %xmm4, %xmm8, %xmm4
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm0, %xmm8
-; AVX2-NEXT: vpsrlw $8, %xmm8, %xmm8
-; AVX2-NEXT: vpand %xmm4, %xmm8, %xmm8
-; AVX2-NEXT: vpsrlw $8, %xmm4, %xmm4
-; AVX2-NEXT: vpaddw %xmm4, %xmm8, %xmm4
-; AVX2-NEXT: vpshufb %xmm1, %xmm5, %xmm8
-; AVX2-NEXT: vpsrlw $4, %xmm1, %xmm9
-; AVX2-NEXT: vpand %xmm7, %xmm9, %xmm9
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm9, %xmm10
-; AVX2-NEXT: vpand %xmm10, %xmm8, %xmm8
-; AVX2-NEXT: vpshufb %xmm9, %xmm5, %xmm9
-; AVX2-NEXT: vpaddb %xmm9, %xmm8, %xmm8
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm1, %xmm9
-; AVX2-NEXT: vpsrlw $8, %xmm9, %xmm9
-; AVX2-NEXT: vpand %xmm9, %xmm8, %xmm9
-; AVX2-NEXT: vpsrlw $8, %xmm8, %xmm8
-; AVX2-NEXT: vpaddw %xmm9, %xmm8, %xmm8
-; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm9
-; AVX2-NEXT: vpsrlw $4, %xmm2, %xmm10
-; AVX2-NEXT: vpand %xmm7, %xmm10, %xmm10
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm10, %xmm11
-; AVX2-NEXT: vpand %xmm11, %xmm9, %xmm9
-; AVX2-NEXT: vpshufb %xmm10, %xmm5, %xmm10
-; AVX2-NEXT: vpaddb %xmm10, %xmm9, %xmm9
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm2, %xmm10
-; AVX2-NEXT: vpsrlw $8, %xmm10, %xmm10
-; AVX2-NEXT: vpand %xmm10, %xmm9, %xmm10
-; AVX2-NEXT: vpsrlw $8, %xmm9, %xmm9
-; AVX2-NEXT: vpaddw %xmm10, %xmm9, %xmm9
-; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm10
-; AVX2-NEXT: vpsrlw $4, %xmm3, %xmm11
-; AVX2-NEXT: vpand %xmm7, %xmm11, %xmm7
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm7, %xmm11
-; AVX2-NEXT: vpand %xmm11, %xmm10, %xmm10
-; AVX2-NEXT: vpshufb %xmm7, %xmm5, %xmm5
-; AVX2-NEXT: vpaddb %xmm5, %xmm10, %xmm5
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm3, %xmm6
-; AVX2-NEXT: vpsrlw $8, %xmm6, %xmm6
-; AVX2-NEXT: vpand %xmm6, %xmm5, %xmm6
-; AVX2-NEXT: vpsrlw $8, %xmm5, %xmm5
-; AVX2-NEXT: vpaddw %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
-; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3
-; AVX2-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX2-NEXT: vpshufb %ymm1, %ymm3, %ymm4
+; AVX2-NEXT: vpsrlw $4, %ymm1, %ymm5
+; AVX2-NEXT: vpbroadcastb {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; AVX2-NEXT: vpand %ymm6, %ymm5, %ymm5
+; AVX2-NEXT: vpxor %xmm7, %xmm7, %xmm7
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm5, %ymm8
+; AVX2-NEXT: vpand %ymm4, %ymm8, %ymm4
+; AVX2-NEXT: vpshufb %ymm5, %ymm3, %ymm5
+; AVX2-NEXT: vpaddb %ymm5, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm1, %ymm5
+; AVX2-NEXT: vpsrlw $8, %ymm5, %ymm5
+; AVX2-NEXT: vpand %ymm5, %ymm4, %ymm5
+; AVX2-NEXT: vpsrlw $8, %ymm4, %ymm4
+; AVX2-NEXT: vpaddw %ymm5, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqw %ymm7, %ymm1, %ymm1
; AVX2-NEXT: vpsrld $16, %ymm1, %ymm1
-; AVX2-NEXT: vpand %ymm1, %ymm5, %ymm1
-; AVX2-NEXT: vpsrld $16, %ymm5, %ymm5
-; AVX2-NEXT: vpaddd %ymm1, %ymm5, %ymm1
-; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm4
+; AVX2-NEXT: vpand %ymm1, %ymm4, %ymm1
+; AVX2-NEXT: vpsrld $16, %ymm4, %ymm4
+; AVX2-NEXT: vpaddd %ymm1, %ymm4, %ymm1
; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
-; AVX2-NEXT: vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm2
+; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm4
+; AVX2-NEXT: vpand %ymm6, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm4, %ymm5
+; AVX2-NEXT: vpand %ymm5, %ymm2, %ymm2
+; AVX2-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX2-NEXT: vpaddb %ymm3, %ymm2, %ymm2
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm0, %ymm3
+; AVX2-NEXT: vpsrlw $8, %ymm3, %ymm3
+; AVX2-NEXT: vpand %ymm3, %ymm2, %ymm3
+; AVX2-NEXT: vpsrlw $8, %ymm2, %ymm2
+; AVX2-NEXT: vpaddw %ymm3, %ymm2, %ymm2
+; AVX2-NEXT: vpcmpeqw %ymm7, %ymm0, %ymm0
; AVX2-NEXT: vpsrld $16, %ymm0, %ymm0
-; AVX2-NEXT: vpand %ymm0, %ymm4, %ymm0
-; AVX2-NEXT: vpsrld $16, %ymm4, %ymm2
+; AVX2-NEXT: vpand %ymm0, %ymm2, %ymm0
+; AVX2-NEXT: vpsrld $16, %ymm2, %ymm2
; AVX2-NEXT: vpaddd %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT: retq
@@ -1008,71 +982,45 @@ define <8 x i32> @widen_ctlz_undef_v2i32_v8i32(<2 x i32> %a0, <2 x i32> %a1, <2
; AVX2: # %bb.0:
; AVX2-NEXT: # kill: def $xmm1 killed $xmm1 def $ymm1
; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX2-NEXT: vmovq {{.*#+}} xmm5 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
-; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm4
-; AVX2-NEXT: vpsrlw $4, %xmm0, %xmm6
-; AVX2-NEXT: vpbroadcastb {{.*#+}} xmm7 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; AVX2-NEXT: vpand %xmm7, %xmm6, %xmm8
-; AVX2-NEXT: vpxor %xmm6, %xmm6, %xmm6
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm8, %xmm9
-; AVX2-NEXT: vpand %xmm4, %xmm9, %xmm4
-; AVX2-NEXT: vpshufb %xmm8, %xmm5, %xmm8
-; AVX2-NEXT: vpaddb %xmm4, %xmm8, %xmm4
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm0, %xmm8
-; AVX2-NEXT: vpsrlw $8, %xmm8, %xmm8
-; AVX2-NEXT: vpand %xmm4, %xmm8, %xmm8
-; AVX2-NEXT: vpsrlw $8, %xmm4, %xmm4
-; AVX2-NEXT: vpaddw %xmm4, %xmm8, %xmm4
-; AVX2-NEXT: vpshufb %xmm1, %xmm5, %xmm8
-; AVX2-NEXT: vpsrlw $4, %xmm1, %xmm9
-; AVX2-NEXT: vpand %xmm7, %xmm9, %xmm9
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm9, %xmm10
-; AVX2-NEXT: vpand %xmm10, %xmm8, %xmm8
-; AVX2-NEXT: vpshufb %xmm9, %xmm5, %xmm9
-; AVX2-NEXT: vpaddb %xmm9, %xmm8, %xmm8
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm1, %xmm9
-; AVX2-NEXT: vpsrlw $8, %xmm9, %xmm9
-; AVX2-NEXT: vpand %xmm9, %xmm8, %xmm9
-; AVX2-NEXT: vpsrlw $8, %xmm8, %xmm8
-; AVX2-NEXT: vpaddw %xmm9, %xmm8, %xmm8
-; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm9
-; AVX2-NEXT: vpsrlw $4, %xmm2, %xmm10
-; AVX2-NEXT: vpand %xmm7, %xmm10, %xmm10
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm10, %xmm11
-; AVX2-NEXT: vpand %xmm11, %xmm9, %xmm9
-; AVX2-NEXT: vpshufb %xmm10, %xmm5, %xmm10
-; AVX2-NEXT: vpaddb %xmm10, %xmm9, %xmm9
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm2, %xmm10
-; AVX2-NEXT: vpsrlw $8, %xmm10, %xmm10
-; AVX2-NEXT: vpand %xmm10, %xmm9, %xmm10
-; AVX2-NEXT: vpsrlw $8, %xmm9, %xmm9
-; AVX2-NEXT: vpaddw %xmm10, %xmm9, %xmm9
-; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm10
-; AVX2-NEXT: vpsrlw $4, %xmm3, %xmm11
-; AVX2-NEXT: vpand %xmm7, %xmm11, %xmm7
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm7, %xmm11
-; AVX2-NEXT: vpand %xmm11, %xmm10, %xmm10
-; AVX2-NEXT: vpshufb %xmm7, %xmm5, %xmm5
-; AVX2-NEXT: vpaddb %xmm5, %xmm10, %xmm5
-; AVX2-NEXT: vpcmpeqb %xmm6, %xmm3, %xmm6
-; AVX2-NEXT: vpsrlw $8, %xmm6, %xmm6
-; AVX2-NEXT: vpand %xmm6, %xmm5, %xmm6
-; AVX2-NEXT: vpsrlw $8, %xmm5, %xmm5
-; AVX2-NEXT: vpaddw %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5
; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
-; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3
-; AVX2-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm1
+; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
+; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX2-NEXT: vpshufb %ymm1, %ymm3, %ymm4
+; AVX2-NEXT: vpsrlw $4, %ymm1, %ymm5
+; AVX2-NEXT: vpbroadcastb {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; AVX2-NEXT: vpand %ymm6, %ymm5, %ymm5
+; AVX2-NEXT: vpxor %xmm7, %xmm7, %xmm7
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm5, %ymm8
+; AVX2-NEXT: vpand %ymm4, %ymm8, %ymm4
+; AVX2-NEXT: vpshufb %ymm5, %ymm3, %ymm5
+; AVX2-NEXT: vpaddb %ymm5, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm1, %ymm5
+; AVX2-NEXT: vpsrlw $8, %ymm5, %ymm5
+; AVX2-NEXT: vpand %ymm5, %ymm4, %ymm5
+; AVX2-NEXT: vpsrlw $8, %ymm4, %ymm4
+; AVX2-NEXT: vpaddw %ymm5, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqw %ymm7, %ymm1, %ymm1
; AVX2-NEXT: vpsrld $16, %ymm1, %ymm1
-; AVX2-NEXT: vpand %ymm1, %ymm5, %ymm1
-; AVX2-NEXT: vpsrld $16, %ymm5, %ymm5
-; AVX2-NEXT: vpaddd %ymm1, %ymm5, %ymm1
-; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm4
+; AVX2-NEXT: vpand %ymm1, %ymm4, %ymm1
+; AVX2-NEXT: vpsrld $16, %ymm4, %ymm4
+; AVX2-NEXT: vpaddd %ymm1, %ymm4, %ymm1
; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
-; AVX2-NEXT: vpcmpeqw %ymm3, %ymm0, %ymm0
+; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm2
+; AVX2-NEXT: vpsrlw $4, %ymm0, %ymm4
+; AVX2-NEXT: vpand %ymm6, %ymm4, %ymm4
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm4, %ymm5
+; AVX2-NEXT: vpand %ymm5, %ymm2, %ymm2
+; AVX2-NEXT: vpshufb %ymm4, %ymm3, %ymm3
+; AVX2-NEXT: vpaddb %ymm3, %ymm2, %ymm2
+; AVX2-NEXT: vpcmpeqb %ymm7, %ymm0, %ymm3
+; AVX2-NEXT: vpsrlw $8, %ymm3, %ymm3
+; AVX2-NEXT: vpand %ymm3, %ymm2, %ymm3
+; AVX2-NEXT: vpsrlw $8, %ymm2, %ymm2
+; AVX2-NEXT: vpaddw %ymm3, %ymm2, %ymm2
+; AVX2-NEXT: vpcmpeqw %ymm7, %ymm0, %ymm0
; AVX2-NEXT: vpsrld $16, %ymm0, %ymm0
-; AVX2-NEXT: vpand %ymm0, %ymm4, %ymm0
-; AVX2-NEXT: vpsrld $16, %ymm4, %ymm2
+; AVX2-NEXT: vpand %ymm0, %ymm2, %ymm0
+; AVX2-NEXT: vpsrld $16, %ymm2, %ymm2
; AVX2-NEXT: vpaddd %ymm0, %ymm2, %ymm0
; AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
; AVX2-NEXT: retq
More information about the llvm-commits
mailing list