[llvm] 170ba6e - [X86] combineINSERT_SUBVECTOR - attempt to combine concatenated shuffles
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Wed Sep 20 04:31:15 PDT 2023
Author: Simon Pilgrim
Date: 2023-09-20T12:17:51+01:00
New Revision: 170ba6ee12182b8f500ffad1c28d3a3d8de6368d
URL: https://github.com/llvm/llvm-project/commit/170ba6ee12182b8f500ffad1c28d3a3d8de6368d
DIFF: https://github.com/llvm/llvm-project/commit/170ba6ee12182b8f500ffad1c28d3a3d8de6368d.diff
LOG: [X86] combineINSERT_SUBVECTOR - attempt to combine concatenated shuffles
If all the concatenated subvectors are targets shuffle nodes, then call combineX86ShufflesRecursively to attempt to combine them.
Unlike the existing shuffle concatenation in collectConcatOps, this isn't limited to splat cases and won't attempt to concat the source nodes prior to creating the larger shuffle node, so will usually only combine to create cross-lane shuffles.
This exposed a hidden issue in matchBinaryShuffle that wasn't limiting v64i8/v32i16 UNPACK nodes to AVX512BW targets.
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.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-i64-stride-4.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
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5c0c3af364259c1..07248794acc4c2f 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37317,7 +37317,8 @@ static bool matchBinaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
(MaskVT.is128BitVector() && Subtarget.hasSSE2()) ||
(MaskVT.is256BitVector() && 32 <= EltSizeInBits && Subtarget.hasAVX()) ||
(MaskVT.is256BitVector() && Subtarget.hasAVX2()) ||
- (MaskVT.is512BitVector() && Subtarget.hasAVX512())) {
+ (MaskVT.is512BitVector() && Subtarget.hasAVX512() &&
+ (32 <= EltSizeInBits || Subtarget.hasBWI()))) {
if (matchShuffleWithUNPCK(MaskVT, V1, V2, Shuffle, IsUnary, Mask, DL, DAG,
Subtarget)) {
SrcVT = DstVT = MaskVT;
@@ -55005,6 +55006,15 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT,
getZeroVector(OpVT, Subtarget, DAG, dl),
SubVectorOps[0], DAG.getIntPtrConstant(0, dl));
+
+ // Attempt to recursively combine to a shuffle.
+ if (all_of(SubVectorOps, [](SDValue SubOp) {
+ return isTargetShuffle(SubOp.getOpcode());
+ })) {
+ SDValue Op(N, 0);
+ if (SDValue Res = combineX86ShufflesRecursively(Op, DAG, Subtarget))
+ return Res;
+ }
}
// If this is a broadcast insert into an upper undef, use a larger broadcast.
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
index 0e3ba8789ed1b8c..74cd79ab551e2a7 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
@@ -433,58 +433,157 @@ define void @store_i16_stride4_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
-; AVX2-ONLY-LABEL: store_i16_stride4_vf16:
-; AVX2-ONLY: # %bb.0:
-; AVX2-ONLY-NEXT: vmovdqa (%rcx), %xmm5
-; AVX2-ONLY-NEXT: vmovdqa 16(%rcx), %xmm0
-; AVX2-ONLY-NEXT: vmovdqa (%rdx), %xmm6
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdx), %xmm1
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX2-ONLY-NEXT: vmovdqa (%rsi), %xmm7
-; AVX2-ONLY-NEXT: vmovdqa 16(%rsi), %xmm3
-; AVX2-ONLY-NEXT: vmovdqa (%rdi), %xmm8
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdi), %xmm4
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm2 = ymm9[0],ymm2[1],ymm9[2],ymm2[3],ymm9[4],ymm2[5],ymm9[6],ymm2[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
-; AVX2-ONLY-NEXT: vmovdqa %ymm0, 64(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm6, 96(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm5, 32(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm2, (%r8)
-; AVX2-ONLY-NEXT: vzeroupper
-; AVX2-ONLY-NEXT: retq
+; AVX2-SLOW-LABEL: store_i16_stride4_vf16:
+; AVX2-SLOW: # %bb.0:
+; AVX2-SLOW-NEXT: vmovdqa (%rcx), %xmm5
+; AVX2-SLOW-NEXT: vmovdqa 16(%rcx), %xmm0
+; AVX2-SLOW-NEXT: vmovdqa (%rdx), %xmm6
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdx), %xmm1
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-SLOW-NEXT: vmovdqa (%rsi), %xmm7
+; AVX2-SLOW-NEXT: vmovdqa 16(%rsi), %xmm3
+; AVX2-SLOW-NEXT: vmovdqa (%rdi), %xmm8
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdi), %xmm4
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm9[0],ymm2[1],ymm9[2],ymm2[3],ymm9[4],ymm2[5],ymm9[6],ymm2[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-SLOW-NEXT: vmovdqa %ymm0, 64(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm6, 96(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm5, 32(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm2, (%r8)
+; AVX2-SLOW-NEXT: vzeroupper
+; AVX2-SLOW-NEXT: retq
+;
+; AVX2-FAST-LABEL: store_i16_stride4_vf16:
+; AVX2-FAST: # %bb.0:
+; AVX2-FAST-NEXT: vmovdqa (%rsi), %xmm2
+; AVX2-FAST-NEXT: vmovdqa 16(%rsi), %xmm0
+; AVX2-FAST-NEXT: vmovdqa (%rdi), %xmm3
+; AVX2-FAST-NEXT: vmovdqa 16(%rdi), %xmm1
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-FAST-NEXT: vmovdqa (%rcx), %xmm5
+; AVX2-FAST-NEXT: vmovdqa 16(%rcx), %xmm6
+; AVX2-FAST-NEXT: vmovdqa (%rdx), %xmm7
+; AVX2-FAST-NEXT: vmovdqa 16(%rdx), %xmm8
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [0,0,1,1,2,2,3,3]
+; AVX2-FAST-NEXT: vpermd %ymm9, %ymm10, %ymm9
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0],ymm9[1],ymm4[2],ymm9[3],ymm4[4],ymm9[5],ymm4[6],ymm9[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-FAST-NEXT: vpermd %ymm3, %ymm10, %ymm3
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm3[0],zero,xmm3[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX2-FAST-NEXT: vpermd %ymm5, %ymm10, %ymm5
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2],ymm5[3],ymm3[4],ymm5[5],ymm3[6],ymm5[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX2-FAST-NEXT: vpermd %ymm1, %ymm10, %ymm1
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
+; AVX2-FAST-NEXT: vmovdqa %ymm0, 64(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm3, 96(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm2, 32(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm4, (%r8)
+; AVX2-FAST-NEXT: vzeroupper
+; AVX2-FAST-NEXT: retq
+;
+; AVX2-FAST-PERLANE-LABEL: store_i16_stride4_vf16:
+; AVX2-FAST-PERLANE: # %bb.0:
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %xmm5
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rcx), %xmm0
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %xmm6
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdx), %xmm1
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %xmm7
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rsi), %xmm3
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %xmm8
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdi), %xmm4
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm2 = ymm9[0],ymm2[1],ymm9[2],ymm2[3],ymm9[4],ymm2[5],ymm9[6],ymm2[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm0, 64(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm6, 96(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm5, 32(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm2, (%r8)
+; AVX2-FAST-PERLANE-NEXT: vzeroupper
+; AVX2-FAST-PERLANE-NEXT: retq
;
; AVX512F-LABEL: store_i16_stride4_vf16:
; AVX512F: # %bb.0:
@@ -734,106 +833,293 @@ define void @store_i16_stride4_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
-; AVX2-ONLY-LABEL: store_i16_stride4_vf32:
-; AVX2-ONLY: # %bb.0:
-; AVX2-ONLY-NEXT: vmovdqa (%rcx), %xmm3
-; AVX2-ONLY-NEXT: vmovdqa 16(%rcx), %xmm0
-; AVX2-ONLY-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX2-ONLY-NEXT: vmovdqa 48(%rcx), %xmm7
-; AVX2-ONLY-NEXT: vmovdqa (%rdx), %xmm6
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdx), %xmm1
-; AVX2-ONLY-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX2-ONLY-NEXT: vmovdqa 48(%rdx), %xmm8
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm2[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm2, %ymm9, %ymm2
-; AVX2-ONLY-NEXT: vmovdqa (%rsi), %xmm9
-; AVX2-ONLY-NEXT: vmovdqa (%rdi), %xmm10
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
-; AVX2-ONLY-NEXT: vmovdqa 48(%rsi), %xmm12
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7]
-; AVX2-ONLY-NEXT: vmovdqa 48(%rdi), %xmm11
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm3, %ymm6, %ymm3
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm6[0],zero,xmm6[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2],ymm3[3],ymm6[4],ymm3[5],ymm6[6],ymm3[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX2-ONLY-NEXT: vmovdqa 32(%rsi), %xmm10
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1],ymm9[2],ymm6[3],ymm9[4],ymm6[5],ymm9[6],ymm6[7]
-; AVX2-ONLY-NEXT: vmovdqa 32(%rdi), %xmm9
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm7[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2],ymm7[3],ymm8[4],ymm7[5],ymm8[6],ymm7[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm8[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
-; AVX2-ONLY-NEXT: vmovdqa 16(%rsi), %xmm12
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm8 = ymm11[0],ymm8[1],ymm11[2],ymm8[3],ymm11[4],ymm8[5],ymm11[6],ymm8[7]
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdi), %xmm11
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm5[0],zero,xmm5[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4],ymm4[5],ymm5[6],ymm4[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm9, %ymm1
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
-; AVX2-ONLY-NEXT: vmovdqa %ymm0, 96(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm5, 64(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm4, 160(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm8, 128(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm7, 224(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm6, 192(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm3, (%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm2, 32(%r8)
-; AVX2-ONLY-NEXT: vzeroupper
-; AVX2-ONLY-NEXT: retq
+; AVX2-SLOW-LABEL: store_i16_stride4_vf32:
+; AVX2-SLOW: # %bb.0:
+; AVX2-SLOW-NEXT: vmovdqa (%rcx), %xmm3
+; AVX2-SLOW-NEXT: vmovdqa 16(%rcx), %xmm0
+; AVX2-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX2-SLOW-NEXT: vmovdqa 48(%rcx), %xmm7
+; AVX2-SLOW-NEXT: vmovdqa (%rdx), %xmm6
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdx), %xmm1
+; AVX2-SLOW-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX2-SLOW-NEXT: vmovdqa 48(%rdx), %xmm8
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm2[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX2-SLOW-NEXT: vmovdqa (%rsi), %xmm9
+; AVX2-SLOW-NEXT: vmovdqa (%rdi), %xmm10
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-SLOW-NEXT: vmovdqa 48(%rsi), %xmm12
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7]
+; AVX2-SLOW-NEXT: vmovdqa 48(%rdi), %xmm11
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm6, %ymm3
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm6[0],zero,xmm6[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2],ymm3[3],ymm6[4],ymm3[5],ymm6[6],ymm3[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-SLOW-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1],ymm9[2],ymm6[3],ymm9[4],ymm6[5],ymm9[6],ymm6[7]
+; AVX2-SLOW-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm7[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2],ymm7[3],ymm8[4],ymm7[5],ymm8[6],ymm7[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm8[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-SLOW-NEXT: vmovdqa 16(%rsi), %xmm12
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm8 = ymm11[0],ymm8[1],ymm11[2],ymm8[3],ymm11[4],ymm8[5],ymm11[6],ymm8[7]
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdi), %xmm11
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm5[0],zero,xmm5[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4],ymm4[5],ymm5[6],ymm4[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm9, %ymm1
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-SLOW-NEXT: vmovdqa %ymm0, 96(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm5, 64(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm4, 160(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm8, 128(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm7, 224(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm6, 192(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm3, (%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm2, 32(%r8)
+; AVX2-SLOW-NEXT: vzeroupper
+; AVX2-SLOW-NEXT: retq
+;
+; AVX2-FAST-LABEL: store_i16_stride4_vf32:
+; AVX2-FAST: # %bb.0:
+; AVX2-FAST-NEXT: vmovdqa (%rsi), %xmm5
+; AVX2-FAST-NEXT: vmovdqa 16(%rsi), %xmm0
+; AVX2-FAST-NEXT: vmovdqa 32(%rsi), %xmm4
+; AVX2-FAST-NEXT: vmovdqa 48(%rsi), %xmm8
+; AVX2-FAST-NEXT: vmovdqa (%rdi), %xmm7
+; AVX2-FAST-NEXT: vmovdqa 16(%rdi), %xmm1
+; AVX2-FAST-NEXT: vmovdqa 32(%rdi), %xmm6
+; AVX2-FAST-NEXT: vmovdqa 48(%rdi), %xmm9
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-FAST-NEXT: vmovdqa (%rcx), %xmm10
+; AVX2-FAST-NEXT: vmovdqa (%rdx), %xmm11
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [0,0,1,1,2,2,3,3]
+; AVX2-FAST-NEXT: vpermd %ymm12, %ymm3, %ymm12
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm12[1],ymm2[2],ymm12[3],ymm2[4],ymm12[5],ymm2[6],ymm12[7]
+; AVX2-FAST-NEXT: vmovdqa 48(%rcx), %xmm12
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
+; AVX2-FAST-NEXT: vmovdqa 48(%rdx), %xmm13
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX2-FAST-NEXT: vpermd %ymm7, %ymm3, %ymm7
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[3],ymm5[4],ymm7[5],ymm5[6],ymm7[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm7[0],zero,xmm7[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm7, %ymm10, %ymm7
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
+; AVX2-FAST-NEXT: vpermd %ymm10, %ymm3, %ymm10
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0],ymm10[1],ymm7[2],ymm10[3],ymm7[4],ymm10[5],ymm7[6],ymm10[7]
+; AVX2-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm8[0],zero,xmm8[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
+; AVX2-FAST-NEXT: vmovdqa 32(%rdx), %xmm11
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX2-FAST-NEXT: vpermd %ymm9, %ymm3, %ymm9
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0],ymm9[1],ymm8[2],ymm9[3],ymm8[4],ymm9[5],ymm8[6],ymm9[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm9[0],zero,xmm9[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm9, %ymm12, %ymm9
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX2-FAST-NEXT: vpermd %ymm12, %ymm3, %ymm12
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0],ymm12[1],ymm9[2],ymm12[3],ymm9[4],ymm12[5],ymm9[6],ymm12[7]
+; AVX2-FAST-NEXT: vmovdqa 16(%rcx), %xmm12
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm4[0],zero,xmm4[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm4, %ymm6, %ymm4
+; AVX2-FAST-NEXT: vmovdqa 16(%rdx), %xmm6
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX2-FAST-NEXT: vpermd %ymm10, %ymm3, %ymm10
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0],ymm10[1],ymm4[2],ymm10[3],ymm4[4],ymm10[5],ymm4[6],ymm10[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm6[0],xmm12[0],xmm6[1],xmm12[1],xmm6[2],xmm12[2],xmm6[3],xmm12[3]
+; AVX2-FAST-NEXT: vpermd %ymm11, %ymm3, %ymm11
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2],ymm11[3],ymm10[4],ymm11[5],ymm10[6],ymm11[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm12[4],xmm6[5],xmm12[5],xmm6[6],xmm12[6],xmm6[7],xmm12[7]
+; AVX2-FAST-NEXT: vpermd %ymm1, %ymm3, %ymm1
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
+; AVX2-FAST-NEXT: vmovdqa %ymm0, 96(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm10, 64(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm4, 160(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm9, 128(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm8, 224(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm7, 192(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm5, (%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm2, 32(%r8)
+; AVX2-FAST-NEXT: vzeroupper
+; AVX2-FAST-NEXT: retq
+;
+; AVX2-FAST-PERLANE-LABEL: store_i16_stride4_vf32:
+; AVX2-FAST-PERLANE: # %bb.0:
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %xmm3
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rcx), %xmm0
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rcx), %xmm7
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %xmm6
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdx), %xmm1
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rdx), %xmm8
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm2[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm2, %ymm9, %ymm2
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %xmm9
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %xmm10
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rsi), %xmm12
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rdi), %xmm11
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm3, %ymm6, %ymm3
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm6[0],zero,xmm6[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0],ymm3[1],ymm6[2],ymm3[3],ymm6[4],ymm3[5],ymm6[6],ymm3[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1],ymm9[2],ymm6[3],ymm9[4],ymm6[5],ymm9[6],ymm6[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdi), %xmm9
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm7[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2],ymm7[3],ymm8[4],ymm7[5],ymm8[6],ymm7[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm8[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rsi), %xmm12
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm8 = ymm11[0],ymm8[1],ymm11[2],ymm8[3],ymm11[4],ymm8[5],ymm11[6],ymm8[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdi), %xmm11
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm5[0],zero,xmm5[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4],ymm4[5],ymm5[6],ymm4[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm5, %ymm9, %ymm5
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm1, %ymm9, %ymm1
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm0, 96(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm5, 64(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm4, 160(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm8, 128(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm7, 224(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm6, 192(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm3, (%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm2, 32(%r8)
+; AVX2-FAST-PERLANE-NEXT: vzeroupper
+; AVX2-FAST-PERLANE-NEXT: retq
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride4_vf32:
; AVX512F-ONLY-SLOW: # %bb.0:
@@ -938,108 +1224,85 @@ define void @store_i16_stride4_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-ONLY-SLOW-NEXT: vzeroupper
; AVX512F-ONLY-SLOW-NEXT: retq
;
-; AVX512F-ONLY-FAST-LABEL: store_i16_stride4_vf32:
-; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rcx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rcx), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rdx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm7, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rsi), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rsi), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rdi), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rdi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm4[4],xmm14[5],xmm4[5],xmm14[6],xmm4[6],xmm14[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm4[0],zero,xmm4[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm14, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm4, %zmm4
-; AVX512F-ONLY-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm13, %zmm4 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm9, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm10[0],zero,xmm10[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm10, %ymm13, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm9, %zmm10 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm9[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm9, %ymm11, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm5, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm6[0],zero,xmm6[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm6, %ymm9, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm7, %ymm8, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm1[0],zero,xmm1[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm1, %ymm5, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, (%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 192(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 128(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 64(%r8)
-; AVX512F-ONLY-FAST-NEXT: vzeroupper
-; AVX512F-ONLY-FAST-NEXT: retq
+; AVX512F-FAST-LABEL: store_i16_stride4_vf32:
+; AVX512F-FAST: # %bb.0:
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 16(%rsi), %xmm1
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa 48(%rsi), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa 16(%rdi), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm10
+; AVX512F-FAST-NEXT: vmovdqa 48(%rdi), %xmm6
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm4, %ymm7, %ymm4
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 16(%rcx), %xmm11
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
+; AVX512F-FAST-NEXT: vmovdqa 48(%rcx), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa 16(%rdx), %xmm13
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm14
+; AVX512F-FAST-NEXT: vmovdqa 48(%rdx), %xmm8
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm15, %zmm11, %zmm11
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vpermd %zmm11, %zmm13, %zmm1 {%k1}
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm11[0],zero,xmm11[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm11, %ymm15, %ymm11
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm9, %zmm9
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm10, %zmm11, %zmm10
+; AVX512F-FAST-NEXT: vpermd %zmm10, %zmm13, %zmm9 {%k1}
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm6, %zmm7, %zmm6
+; AVX512F-FAST-NEXT: vpermd %zmm6, %zmm13, %zmm5 {%k1}
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2
+; AVX512F-FAST-NEXT: vpermd %zmm2, %zmm13, %zmm0 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, 192(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, 128(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 64(%r8)
+; AVX512F-FAST-NEXT: vzeroupper
+; AVX512F-FAST-NEXT: retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride4_vf32:
; AVX512DQ-SLOW: # %bb.0:
@@ -1140,105 +1403,6 @@ define void @store_i16_stride4_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-SLOW-NEXT: vzeroupper
; AVX512DQ-SLOW-NEXT: retq
;
-; AVX512DQ-FAST-LABEL: store_i16_stride4_vf32:
-; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rdi), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm4[0],zero,xmm4[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm4, %ymm8, %ymm4
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rcx), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm11
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rcx), %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rdx), %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm12
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rdx), %xmm9
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm15[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
-; AVX512DQ-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm13, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm13[0],zero,xmm13[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm7[0],zero,xmm7[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm7, %ymm10, %ymm7
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm10[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm10, %ymm13, %ymm10
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm12 = xmm11[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm11, %zmm7 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm6[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm10, %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm8[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm8, %zmm5 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm6[0],zero,xmm6[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm2, %ymm6, %ymm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, (%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 192(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 128(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, 64(%r8)
-; AVX512DQ-FAST-NEXT: vzeroupper
-; AVX512DQ-FAST-NEXT: retq
-;
; AVX512BW-LABEL: store_i16_stride4_vf32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0
@@ -1699,214 +1863,601 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
-; AVX2-ONLY-LABEL: store_i16_stride4_vf64:
-; AVX2-ONLY: # %bb.0:
-; AVX2-ONLY-NEXT: subq $40, %rsp
-; AVX2-ONLY-NEXT: vmovdqa 16(%rcx), %xmm1
-; AVX2-ONLY-NEXT: vmovdqa 32(%rcx), %xmm9
-; AVX2-ONLY-NEXT: vmovdqa 48(%rcx), %xmm6
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdx), %xmm2
-; AVX2-ONLY-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX2-ONLY-NEXT: vmovdqa 48(%rdx), %xmm8
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX2-ONLY-NEXT: vmovdqa 16(%rsi), %xmm3
-; AVX2-ONLY-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX2-ONLY-NEXT: vmovdqa 16(%rdi), %xmm4
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
-; AVX2-ONLY-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
-; AVX2-ONLY-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-ONLY-NEXT: vmovdqa 64(%rcx), %xmm3
-; AVX2-ONLY-NEXT: vmovdqa 64(%rdx), %xmm4
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
-; AVX2-ONLY-NEXT: vmovdqa 64(%rsi), %xmm5
-; AVX2-ONLY-NEXT: vmovdqa 64(%rdi), %xmm7
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm12[0],zero,xmm12[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm12, %ymm13, %ymm12
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0],ymm2[1],ymm12[2],ymm2[3],ymm12[4],ymm2[5],ymm12[6],ymm2[7]
-; AVX2-ONLY-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
-; AVX2-ONLY-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-ONLY-NEXT: vmovdqa 80(%rcx), %xmm5
-; AVX2-ONLY-NEXT: vmovdqa 80(%rdx), %xmm7
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm12 = xmm4[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm4, %ymm12, %ymm4
-; AVX2-ONLY-NEXT: vmovdqa 80(%rsi), %xmm13
-; AVX2-ONLY-NEXT: vmovdqa 80(%rdi), %xmm14
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm12[0],zero,xmm12[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm12, %ymm15, %ymm15
-; AVX2-ONLY-NEXT: vmovdqa 48(%rsi), %xmm12
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0],ymm4[1],ymm15[2],ymm4[3],ymm15[4],ymm4[5],ymm15[6],ymm4[7]
-; AVX2-ONLY-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX2-ONLY-NEXT: vmovdqa 32(%rdi), %xmm15
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm7[0],zero,xmm7[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2],ymm5[3],ymm7[4],ymm5[5],ymm7[6],ymm5[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm7[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm13[0],zero,xmm13[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
-; AVX2-ONLY-NEXT: vmovdqa 48(%rdi), %xmm14
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm7 = ymm13[0],ymm7[1],ymm13[2],ymm7[3],ymm13[4],ymm7[5],ymm13[6],ymm7[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2],ymm9[3],ymm10[4],ymm9[5],ymm10[6],ymm9[7]
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm10[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm11[0],zero,xmm11[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm11, %ymm13, %ymm11
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0],ymm10[1],ymm11[2],ymm10[3],ymm11[4],ymm10[5],ymm11[6],ymm10[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm6
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
-; AVX2-ONLY-NEXT: vmovdqa 96(%rcx), %xmm11
-; AVX2-ONLY-NEXT: vmovdqa 96(%rdx), %xmm12
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm8[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm8, %ymm13, %ymm8
-; AVX2-ONLY-NEXT: vmovdqa 96(%rsi), %xmm13
-; AVX2-ONLY-NEXT: vmovdqa 96(%rdi), %xmm14
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm15[0],zero,xmm15[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm15, %ymm0, %ymm0
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0],ymm0[1],ymm11[2],ymm0[3],ymm11[4],ymm0[5],ymm11[6],ymm0[7]
-; AVX2-ONLY-NEXT: vmovdqa 112(%rcx), %xmm1
-; AVX2-ONLY-NEXT: vmovdqa 112(%rdx), %xmm13
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm14 = xmm12[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm12, %ymm14, %ymm12
-; AVX2-ONLY-NEXT: vmovdqa 112(%rsi), %xmm14
-; AVX2-ONLY-NEXT: vmovdqa 112(%rdi), %xmm15
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0],ymm12[1],ymm0[2],ymm12[3],ymm0[4],ymm12[5],ymm0[6],ymm12[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX2-ONLY-NEXT: vmovdqa (%rcx), %xmm2
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm4 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
-; AVX2-ONLY-NEXT: vmovdqa (%rdx), %xmm1
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
-; AVX2-ONLY-NEXT: vmovdqa (%rsi), %xmm14
-; AVX2-ONLY-NEXT: vmovdqa (%rdi), %xmm15
-; AVX2-ONLY-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2],ymm13[3],ymm0[4],ymm13[5],ymm0[6],ymm13[7]
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX2-ONLY-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
-; AVX2-ONLY-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
-; AVX2-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX2-ONLY-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX2-ONLY-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
-; AVX2-ONLY-NEXT: vmovdqa %ymm1, 32(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm0, (%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm4, 480(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm12, 448(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm11, 416(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm8, 384(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm6, 224(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm10, 192(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm9, 160(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm7, 128(%r8)
-; AVX2-ONLY-NEXT: vmovdqa %ymm5, 352(%r8)
-; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-ONLY-NEXT: vmovaps %ymm0, 320(%r8)
-; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-ONLY-NEXT: vmovaps %ymm0, 288(%r8)
-; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-ONLY-NEXT: vmovaps %ymm0, 256(%r8)
-; AVX2-ONLY-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX2-ONLY-NEXT: vmovaps %ymm0, 64(%r8)
-; AVX2-ONLY-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
-; AVX2-ONLY-NEXT: vmovaps %ymm0, 96(%r8)
-; AVX2-ONLY-NEXT: addq $40, %rsp
-; AVX2-ONLY-NEXT: vzeroupper
-; AVX2-ONLY-NEXT: retq
+; AVX2-SLOW-LABEL: store_i16_stride4_vf64:
+; AVX2-SLOW: # %bb.0:
+; AVX2-SLOW-NEXT: subq $40, %rsp
+; AVX2-SLOW-NEXT: vmovdqa 16(%rcx), %xmm1
+; AVX2-SLOW-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX2-SLOW-NEXT: vmovdqa 48(%rcx), %xmm6
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdx), %xmm2
+; AVX2-SLOW-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX2-SLOW-NEXT: vmovdqa 48(%rdx), %xmm8
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX2-SLOW-NEXT: vmovdqa 16(%rsi), %xmm3
+; AVX2-SLOW-NEXT: vmovdqa 32(%rsi), %xmm11
+; AVX2-SLOW-NEXT: vmovdqa 16(%rdi), %xmm4
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
+; AVX2-SLOW-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
+; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-SLOW-NEXT: vmovdqa 64(%rcx), %xmm3
+; AVX2-SLOW-NEXT: vmovdqa 64(%rdx), %xmm4
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX2-SLOW-NEXT: vmovdqa 64(%rsi), %xmm5
+; AVX2-SLOW-NEXT: vmovdqa 64(%rdi), %xmm7
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm12[0],zero,xmm12[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm13, %ymm12
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0],ymm2[1],ymm12[2],ymm2[3],ymm12[4],ymm2[5],ymm12[6],ymm2[7]
+; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
+; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-SLOW-NEXT: vmovdqa 80(%rcx), %xmm5
+; AVX2-SLOW-NEXT: vmovdqa 80(%rdx), %xmm7
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm4[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm12, %ymm4
+; AVX2-SLOW-NEXT: vmovdqa 80(%rsi), %xmm13
+; AVX2-SLOW-NEXT: vmovdqa 80(%rdi), %xmm14
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm12[0],zero,xmm12[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm15, %ymm15
+; AVX2-SLOW-NEXT: vmovdqa 48(%rsi), %xmm12
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0],ymm4[1],ymm15[2],ymm4[3],ymm15[4],ymm4[5],ymm15[6],ymm4[7]
+; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-SLOW-NEXT: vmovdqa 32(%rdi), %xmm15
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm7[0],zero,xmm7[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2],ymm5[3],ymm7[4],ymm5[5],ymm7[6],ymm5[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm7[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm13[0],zero,xmm13[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
+; AVX2-SLOW-NEXT: vmovdqa 48(%rdi), %xmm14
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm7 = ymm13[0],ymm7[1],ymm13[2],ymm7[3],ymm13[4],ymm7[5],ymm13[6],ymm7[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2],ymm9[3],ymm10[4],ymm9[5],ymm10[6],ymm9[7]
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm10[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm11[0],zero,xmm11[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm13, %ymm11
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0],ymm10[1],ymm11[2],ymm10[3],ymm11[4],ymm10[5],ymm11[6],ymm10[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm6
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
+; AVX2-SLOW-NEXT: vmovdqa 96(%rcx), %xmm11
+; AVX2-SLOW-NEXT: vmovdqa 96(%rdx), %xmm12
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm8[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm8, %ymm13, %ymm8
+; AVX2-SLOW-NEXT: vmovdqa 96(%rsi), %xmm13
+; AVX2-SLOW-NEXT: vmovdqa 96(%rdi), %xmm14
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm15[0],zero,xmm15[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm15, %ymm0, %ymm0
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0],ymm0[1],ymm11[2],ymm0[3],ymm11[4],ymm0[5],ymm11[6],ymm0[7]
+; AVX2-SLOW-NEXT: vmovdqa 112(%rcx), %xmm1
+; AVX2-SLOW-NEXT: vmovdqa 112(%rdx), %xmm13
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm12[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm12, %ymm14, %ymm12
+; AVX2-SLOW-NEXT: vmovdqa 112(%rsi), %xmm14
+; AVX2-SLOW-NEXT: vmovdqa 112(%rdi), %xmm15
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0],ymm12[1],ymm0[2],ymm12[3],ymm0[4],ymm12[5],ymm0[6],ymm12[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-SLOW-NEXT: vmovdqa (%rcx), %xmm2
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-SLOW-NEXT: vmovdqa (%rdx), %xmm1
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
+; AVX2-SLOW-NEXT: vmovdqa (%rsi), %xmm14
+; AVX2-SLOW-NEXT: vmovdqa (%rdi), %xmm15
+; AVX2-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2],ymm13[3],ymm0[4],ymm13[5],ymm0[6],ymm13[7]
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX2-SLOW-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
+; AVX2-SLOW-NEXT: vmovdqa %ymm1, 32(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm0, (%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm4, 480(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm12, 448(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm11, 416(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm8, 384(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm6, 224(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm10, 192(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm9, 160(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm7, 128(%r8)
+; AVX2-SLOW-NEXT: vmovdqa %ymm5, 352(%r8)
+; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-SLOW-NEXT: vmovaps %ymm0, 320(%r8)
+; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-SLOW-NEXT: vmovaps %ymm0, 288(%r8)
+; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-SLOW-NEXT: vmovaps %ymm0, 256(%r8)
+; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-SLOW-NEXT: vmovaps %ymm0, 64(%r8)
+; AVX2-SLOW-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-SLOW-NEXT: vmovaps %ymm0, 96(%r8)
+; AVX2-SLOW-NEXT: addq $40, %rsp
+; AVX2-SLOW-NEXT: vzeroupper
+; AVX2-SLOW-NEXT: retq
+;
+; AVX2-FAST-LABEL: store_i16_stride4_vf64:
+; AVX2-FAST: # %bb.0:
+; AVX2-FAST-NEXT: subq $40, %rsp
+; AVX2-FAST-NEXT: vmovdqa 16(%rsi), %xmm1
+; AVX2-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX2-FAST-NEXT: vmovdqa 48(%rsi), %xmm7
+; AVX2-FAST-NEXT: vmovdqa 16(%rdi), %xmm3
+; AVX2-FAST-NEXT: vmovdqa 32(%rdi), %xmm11
+; AVX2-FAST-NEXT: vmovdqa 48(%rdi), %xmm9
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; AVX2-FAST-NEXT: vmovdqa 16(%rcx), %xmm4
+; AVX2-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
+; AVX2-FAST-NEXT: vmovdqa 16(%rdx), %xmm5
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,0,1,1,2,2,3,3]
+; AVX2-FAST-NEXT: vpermd %ymm6, %ymm2, %ymm6
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2],ymm6[3],ymm0[4],ymm6[5],ymm0[6],ymm6[7]
+; AVX2-FAST-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX2-FAST-NEXT: vpermd %ymm3, %ymm2, %ymm3
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7]
+; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-NEXT: vmovdqa 64(%rsi), %xmm4
+; AVX2-FAST-NEXT: vmovdqa 64(%rdi), %xmm5
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm3[0],zero,xmm3[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm3, %ymm6, %ymm3
+; AVX2-FAST-NEXT: vmovdqa 64(%rcx), %xmm6
+; AVX2-FAST-NEXT: vmovdqa 64(%rdx), %xmm8
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX2-FAST-NEXT: vpermd %ymm13, %ymm2, %ymm13
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm13[1],ymm3[2],ymm13[3],ymm3[4],ymm13[5],ymm3[6],ymm13[7]
+; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX2-FAST-NEXT: vpermd %ymm5, %ymm2, %ymm5
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4],ymm5[5],ymm4[6],ymm5[7]
+; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-NEXT: vmovdqa 80(%rsi), %xmm6
+; AVX2-FAST-NEXT: vmovdqa 80(%rdi), %xmm8
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm5[0],zero,xmm5[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm5, %ymm13, %ymm5
+; AVX2-FAST-NEXT: vmovdqa 80(%rcx), %xmm14
+; AVX2-FAST-NEXT: vmovdqa 80(%rdx), %xmm15
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX2-FAST-NEXT: vpermd %ymm13, %ymm2, %ymm13
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm13[1],ymm5[2],ymm13[3],ymm5[4],ymm13[5],ymm5[6],ymm13[7]
+; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-NEXT: vmovdqa 48(%rcx), %xmm13
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm6[0],zero,xmm6[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm6
+; AVX2-FAST-NEXT: vmovdqa 32(%rdx), %xmm0
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX2-FAST-NEXT: vpermd %ymm8, %ymm2, %ymm8
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2],ymm8[3],ymm6[4],ymm8[5],ymm6[6],ymm8[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm8[0],zero,xmm8[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm8, %ymm14, %ymm8
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3]
+; AVX2-FAST-NEXT: vpermd %ymm14, %ymm2, %ymm14
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0],ymm14[1],ymm8[2],ymm14[3],ymm8[4],ymm14[5],ymm8[6],ymm14[7]
+; AVX2-FAST-NEXT: vmovdqa 48(%rdx), %xmm14
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm12[4],xmm0[5],xmm12[5],xmm0[6],xmm12[6],xmm0[7],xmm12[7]
+; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0],ymm0[1],ymm10[2],ymm0[3],ymm10[4],ymm0[5],ymm10[6],ymm0[7]
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-FAST-NEXT: vpermd %ymm11, %ymm2, %ymm11
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm11 = ymm0[0],ymm11[1],ymm0[2],ymm11[3],ymm0[4],ymm11[5],ymm0[6],ymm11[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm7, %ymm0
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-FAST-NEXT: vpermd %ymm7, %ymm2, %ymm7
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm7 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7]
+; AVX2-FAST-NEXT: vmovdqa 96(%rsi), %xmm0
+; AVX2-FAST-NEXT: vmovdqa 96(%rdi), %xmm12
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm9[0],zero,xmm9[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm9, %ymm13, %ymm9
+; AVX2-FAST-NEXT: vmovdqa 96(%rcx), %xmm13
+; AVX2-FAST-NEXT: vmovdqa 96(%rdx), %xmm14
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-FAST-NEXT: vpermd %ymm15, %ymm2, %ymm15
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0],ymm15[1],ymm9[2],ymm15[3],ymm9[4],ymm15[5],ymm9[6],ymm15[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm0[4],xmm12[5],xmm0[5],xmm12[6],xmm0[6],xmm12[7],xmm0[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm0, %ymm12, %ymm0
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-FAST-NEXT: vpermd %ymm12, %ymm2, %ymm12
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0],ymm12[1],ymm0[2],ymm12[3],ymm0[4],ymm12[5],ymm0[6],ymm12[7]
+; AVX2-FAST-NEXT: vmovdqa 112(%rsi), %xmm1
+; AVX2-FAST-NEXT: vmovdqa 112(%rdi), %xmm14
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm13[0],zero,xmm13[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm13, %ymm15, %ymm13
+; AVX2-FAST-NEXT: vmovdqa 112(%rcx), %xmm15
+; AVX2-FAST-NEXT: vmovdqa 112(%rdx), %xmm0
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
+; AVX2-FAST-NEXT: vpermd %ymm3, %ymm2, %ymm3
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0],ymm3[1],ymm13[2],ymm3[3],ymm13[4],ymm3[5],ymm13[6],ymm3[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm1[4],xmm14[5],xmm1[5],xmm14[6],xmm1[6],xmm14[7],xmm1[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm4 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-FAST-NEXT: vmovdqa (%rsi), %xmm1
+; AVX2-FAST-NEXT: vmovdqa (%rdi), %xmm3
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm14[0],zero,xmm14[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm14, %ymm15, %ymm14
+; AVX2-FAST-NEXT: vmovdqa (%rcx), %xmm15
+; AVX2-FAST-NEXT: vmovdqa (%rdx), %xmm0
+; AVX2-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
+; AVX2-FAST-NEXT: vpermd %ymm5, %ymm2, %ymm5
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm5 = ymm14[0],ymm5[1],ymm14[2],ymm5[3],ymm14[4],ymm5[5],ymm14[6],ymm5[7]
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX2-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1
+; AVX2-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX2-FAST-NEXT: vpermd %ymm0, %ymm2, %ymm0
+; AVX2-FAST-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-FAST-NEXT: vmovdqa %ymm0, 32(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm5, (%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm4, 480(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm13, 448(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm12, 416(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm9, 384(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm7, 224(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm11, 192(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm10, 160(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm8, 128(%r8)
+; AVX2-FAST-NEXT: vmovdqa %ymm6, 352(%r8)
+; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-NEXT: vmovaps %ymm0, 320(%r8)
+; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-NEXT: vmovaps %ymm0, 288(%r8)
+; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-NEXT: vmovaps %ymm0, 256(%r8)
+; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-NEXT: vmovaps %ymm0, 64(%r8)
+; AVX2-FAST-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-FAST-NEXT: vmovaps %ymm0, 96(%r8)
+; AVX2-FAST-NEXT: addq $40, %rsp
+; AVX2-FAST-NEXT: vzeroupper
+; AVX2-FAST-NEXT: retq
+;
+; AVX2-FAST-PERLANE-LABEL: store_i16_stride4_vf64:
+; AVX2-FAST-PERLANE: # %bb.0:
+; AVX2-FAST-PERLANE-NEXT: subq $40, %rsp
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rcx), %xmm1
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rcx), %xmm6
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdx), %xmm2
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rdx), %xmm8
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rsi), %xmm3
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rsi), %xmm11
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 16(%rdi), %xmm4
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 64(%rcx), %xmm3
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 64(%rdx), %xmm4
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 64(%rsi), %xmm5
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 64(%rdi), %xmm7
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm12[0],zero,xmm12[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm12, %ymm13, %ymm12
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0],ymm2[1],ymm12[2],ymm2[3],ymm12[4],ymm2[5],ymm12[6],ymm2[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 80(%rcx), %xmm5
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 80(%rdx), %xmm7
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm12 = xmm4[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm4, %ymm12, %ymm4
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 80(%rsi), %xmm13
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 80(%rdi), %xmm14
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm12[0],zero,xmm12[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm12, %ymm15, %ymm15
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rsi), %xmm12
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0],ymm4[1],ymm15[2],ymm4[3],ymm15[4],ymm4[5],ymm15[6],ymm4[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdi), %xmm15
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm5, %ymm7, %ymm5
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm7[0],zero,xmm7[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2],ymm5[3],ymm7[4],ymm5[5],ymm7[6],ymm5[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm13 = xmm7[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm7, %ymm13, %ymm7
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm13[0],zero,xmm13[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 48(%rdi), %xmm14
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm7 = ymm13[0],ymm7[1],ymm13[2],ymm7[3],ymm13[4],ymm7[5],ymm13[6],ymm7[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm10[0],zero,xmm10[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2],ymm9[3],ymm10[4],ymm9[5],ymm10[6],ymm9[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm10[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm11[0],zero,xmm11[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm11, %ymm13, %ymm11
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0],ymm10[1],ymm11[2],ymm10[3],ymm11[4],ymm10[5],ymm11[6],ymm10[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm6
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm11 = xmm8[0],zero,xmm8[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm8, %ymm11, %ymm8
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 96(%rcx), %xmm11
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 96(%rdx), %xmm12
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm13 = xmm8[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm8, %ymm13, %ymm8
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 96(%rsi), %xmm13
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 96(%rdi), %xmm14
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm15[0],zero,xmm15[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm15, %ymm0, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm0[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm11[0],zero,xmm11[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0],ymm0[1],ymm11[2],ymm0[3],ymm11[4],ymm0[5],ymm11[6],ymm0[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 112(%rcx), %xmm1
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 112(%rdx), %xmm13
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm14 = xmm12[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm12, %ymm14, %ymm12
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 112(%rsi), %xmm14
+; AVX2-FAST-PERLANE-NEXT: vmovdqa 112(%rdi), %xmm15
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0],ymm12[1],ymm0[2],ymm12[3],ymm0[4],ymm12[5],ymm0[6],ymm12[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm1[4],xmm13[5],xmm1[5],xmm13[6],xmm1[6],xmm13[7],xmm1[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %xmm2
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm4 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %xmm1
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm13, %ymm14, %ymm13
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %xmm14
+; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %xmm15
+; AVX2-FAST-PERLANE-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm13[1],ymm0[2],ymm13[3],ymm0[4],ymm13[5],ymm0[6],ymm13[7]
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-FAST-PERLANE-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX2-FAST-PERLANE-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero
+; AVX2-FAST-PERLANE-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX2-FAST-PERLANE-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
+; AVX2-FAST-PERLANE-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7]
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm1, 32(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm0, (%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm4, 480(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm12, 448(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm11, 416(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm8, 384(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm6, 224(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm10, 192(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm9, 160(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm7, 128(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm5, 352(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 320(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 288(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 256(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 64(%r8)
+; AVX2-FAST-PERLANE-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 96(%r8)
+; AVX2-FAST-PERLANE-NEXT: addq $40, %rsp
+; AVX2-FAST-PERLANE-NEXT: vzeroupper
+; AVX2-FAST-PERLANE-NEXT: retq
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride4_vf64:
; AVX512F-ONLY-SLOW: # %bb.0:
@@ -2108,205 +2659,159 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512F-ONLY-SLOW-NEXT: vzeroupper
; AVX512F-ONLY-SLOW-NEXT: retq
;
-; AVX512F-ONLY-FAST-LABEL: store_i16_stride4_vf64:
-; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rcx), %xmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rcx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rcx), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rdx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm4, %ymm7, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rsi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rsi), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 16(%rdi), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 48(%rdi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm3[4],xmm14[5],xmm3[5],xmm14[6],xmm3[6],xmm14[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm14[0],xmm3[0],xmm14[1],xmm3[1],xmm14[2],xmm3[2],xmm14[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm3[0],zero,xmm3[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm14, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm17
-; AVX512F-ONLY-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm9, %zmm17 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm9[4],xmm14[5],xmm9[5],xmm14[6],xmm9[6],xmm14[7],xmm9[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm15[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm14[0],xmm9[0],xmm14[1],xmm9[1],xmm14[2],xmm9[2],xmm14[3],xmm9[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm9, %ymm14, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm9, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm9[4],xmm15[5],xmm9[5],xmm15[6],xmm9[6],xmm15[7],xmm9[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm0[0],zero,xmm0[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm0, %ymm16, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm9[0],zero,xmm9[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm9, %ymm15, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm18
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm14, %zmm18 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 112(%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 112(%rdx), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm15[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm14, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 112(%rsi), %xmm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 112(%rdi), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm3[0],zero,xmm3[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm3, %ymm16, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm14[0],zero,xmm14[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm14, %ymm15, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm14, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm19 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm15[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm15[4],xmm3[4],xmm15[5],xmm3[5],xmm15[6],xmm3[6],xmm15[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm9[0],zero,xmm9[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm9, %ymm16, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm15[0],xmm3[0],xmm15[1],xmm3[1],xmm15[2],xmm3[2],xmm15[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm3[0],zero,xmm3[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm15, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm15 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 80(%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 80(%rdx), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm9[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm9, %ymm16, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 80(%rsi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 80(%rdi), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm9[4],xmm3[4],xmm9[5],xmm3[5],xmm9[6],xmm3[6],xmm9[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm14[0],zero,xmm14[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti32x4 $1, %xmm14, %ymm16, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm3[0],xmm9[1],xmm3[1],xmm9[2],xmm3[2],xmm9[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm3[0],zero,xmm3[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm9, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm3, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm16 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm3[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm9, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm3[0],zero,xmm3[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm3, %ymm9, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm9[0],zero,xmm9[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm9, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm3 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm9, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm5 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm20, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm1, %ymm6, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm1[0],zero,xmm1[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm1, %ymm6, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm2[0],zero,xmm2[1],zero
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, (%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 192(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 128(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 320(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, 256(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, 448(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 384(%r8)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, 64(%r8)
-; AVX512F-ONLY-FAST-NEXT: vzeroupper
-; AVX512F-ONLY-FAST-NEXT: retq
+; AVX512F-FAST-LABEL: store_i16_stride4_vf64:
+; AVX512F-FAST: # %bb.0:
+; AVX512F-FAST-NEXT: vmovdqa64 (%rsi), %xmm19
+; AVX512F-FAST-NEXT: vmovdqa 16(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
+; AVX512F-FAST-NEXT: vmovdqa 48(%rsi), %xmm6
+; AVX512F-FAST-NEXT: vmovdqa64 (%rdi), %xmm20
+; AVX512F-FAST-NEXT: vmovdqa 16(%rdi), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512F-FAST-NEXT: vmovdqa 48(%rdi), %xmm7
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm18
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 16(%rcx), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm13
+; AVX512F-FAST-NEXT: vmovdqa 48(%rcx), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa 16(%rdx), %xmm8
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm14
+; AVX512F-FAST-NEXT: vmovdqa 48(%rdx), %xmm10
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm5[4],xmm8[5],xmm5[5],xmm8[6],xmm5[6],xmm8[7],xmm5[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm15, %zmm5, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
+; AVX512F-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-FAST-NEXT: vpermd %zmm8, %zmm5, %zmm18 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %xmm8
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm15
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti32x4 $1, %xmm0, %ymm16, %ymm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm8[0],zero,xmm8[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm8, %ymm15, %ymm8
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm15
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm5, %zmm8 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa 112(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 112(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm15
+; AVX512F-FAST-NEXT: vmovdqa 112(%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 112(%rdx), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm5, %zmm15 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm2[0],zero,xmm2[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm16, %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm16
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm5, %zmm16 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa 80(%rsi), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 80(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm17 = xmm2[0],zero,xmm2[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti32x4 $1, %xmm2, %ymm17, %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm17
+; AVX512F-FAST-NEXT: vmovdqa 80(%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 80(%rdx), %xmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpermd %zmm0, %zmm5, %zmm17 {%k1}
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vpermd %zmm1, %zmm5, %zmm0 {%k1}
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm2[0],zero,xmm2[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm2, %ymm6, %ymm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm6, %zmm2
+; AVX512F-FAST-NEXT: vpermd %zmm2, %zmm5, %zmm1 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm19, %xmm7
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm20, %xmm9
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm2[0],zero,xmm2[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm2, %ymm6, %ymm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
+; AVX512F-FAST-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm6[0],zero,xmm6[1],zero
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
+; AVX512F-FAST-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm2
+; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm6, %zmm3, %zmm3
+; AVX512F-FAST-NEXT: vpermd %zmm3, %zmm5, %zmm2 {%k1}
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, (%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 192(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 128(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, 320(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm16, 256(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm15, 448(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 384(%r8)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm18, 64(%r8)
+; AVX512F-FAST-NEXT: vzeroupper
+; AVX512F-FAST-NEXT: retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride4_vf64:
; AVX512DQ-SLOW: # %bb.0:
@@ -2499,197 +3004,6 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-SLOW-NEXT: vzeroupper
; AVX512DQ-SLOW-NEXT: retq
;
-; AVX512DQ-FAST-LABEL: store_i16_stride4_vf64:
-; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rsi), %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rdi), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rdi), %xmm7
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm4[0],zero,xmm4[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm4, %ymm6, %ymm4
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm17
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rcx), %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa 16(%rdx), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa 48(%rdx), %xmm9
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm6[4],xmm13[5],xmm6[5],xmm13[6],xmm6[6],xmm13[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm15[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm6[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm13, %ymm6
-; AVX512DQ-FAST-NEXT: movw $-21846, %ax # imm = 0xAAAA
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm6, %zmm17 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm13
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm6[4],xmm13[5],xmm6[5],xmm13[6],xmm6[6],xmm13[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm6[0],zero,xmm6[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm13, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm18
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm13[4],xmm15[5],xmm13[5],xmm15[6],xmm13[6],xmm15[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm0[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm0, %ymm16, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm13[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm13, %ymm15, %ymm13
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm13, %zmm18 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 112(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 112(%rdi), %xmm13
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm13, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm19
-; AVX512DQ-FAST-NEXT: vmovdqa 112(%rcx), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 112(%rdx), %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm6[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm6, %ymm16, %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm0[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm0, %zmm19 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm15[0],zero,xmm15[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm15, %ymm16, %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm15
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm16 = xmm13[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm13, %ymm16, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm13, %zmm0, %zmm15 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 80(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 80(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm16 = xmm13[0],zero,xmm13[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm13, %ymm16, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm16
-; AVX512DQ-FAST-NEXT: vmovdqa 80(%rcx), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 80(%rdx), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm20 = xmm13[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti32x4 $1, %xmm13, %ymm20, %ymm13
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm13, %zmm0, %zmm16 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm10 = xmm6[0],zero,xmm6[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm10, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm10
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm6[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm11, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm6, %zmm10 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm0[0],zero,xmm0[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm0, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm0
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm6, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm5[0],zero,xmm5[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-FAST-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,3,3]
-; AVX512DQ-FAST-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, (%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, 192(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 128(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, 320(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm15, 256(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, 448(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, 384(%r8)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, 64(%r8)
-; AVX512DQ-FAST-NEXT: vzeroupper
-; AVX512DQ-FAST-NEXT: retq
-;
; AVX512BW-LABEL: store_i16_stride4_vf64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0
@@ -2771,8 +3085,10 @@ define void @store_i16_stride4_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512BW-SLOW: {{.*}}
+; AVX512DQ-FAST: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
+; AVX512F-ONLY-FAST: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
index 2abbb6805cf41c4..d821b370f78a699 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll
@@ -2769,324 +2769,314 @@ define void @store_i16_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-SLOW-LABEL: store_i16_stride5_vf32:
; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: vmovdqa64 (%rdx), %ymm18
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm4
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm18[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm9
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm12
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512F-SLOW-NEXT: vpshufb %xmm13, %xmm0, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 (%rdx), %ymm17
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm5
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,2,3,3,7,6,7,7]
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm1
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm1[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3,4],ymm3[5,6,7,8],ymm2[9],ymm3[10],ymm2[11,12],ymm3[13,14,15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,2]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm9
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm3, %xmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm0, %xmm22
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm3[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm7 = ymm5[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm0[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm8
+; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,2,1,3]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,5,6]
+; AVX512F-SLOW-NEXT: vmovdqa64 (%rdi), %ymm20
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm6
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm7 = ymm6[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4,5],ymm6[6],ymm7[7,8],ymm6[9],ymm7[10],ymm6[11],ymm7[12,13],ymm6[14],ymm7[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,2]
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512F-SLOW-NEXT: vmovdqa64 (%rdi), %xmm24
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,2,1,3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,5,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm7[0],ymm3[1],ymm7[2],ymm3[3],ymm7[4,5],ymm3[6],ymm7[7,8],ymm3[9],ymm7[10],ymm3[11],ymm7[12,13],ymm3[14],ymm7[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,2]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm2[0,1,0,1]
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm16, %zmm6
-; AVX512F-SLOW-NEXT: vpbroadcastq 24(%r8), %ymm2
-; AVX512F-SLOW-NEXT: vpbroadcastq 32(%r8), %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm13
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm13, %ymm6
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512F-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm17
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm17[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5],ymm6[6],ymm7[7,8],ymm6[9],ymm7[10,11],ymm6[12],ymm7[13],ymm6[14],ymm7[15]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm16, %zmm3
+; AVX512F-SLOW-NEXT: vpbroadcastq 24(%r8), %ymm0
+; AVX512F-SLOW-NEXT: vpbroadcastq 32(%r8), %ymm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm19
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm19
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm3
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm3, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512F-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm18
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm18[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm7[0],ymm0[1],ymm7[2,3],ymm0[4],ymm7[5],ymm0[6],ymm7[7,8],ymm0[9],ymm7[10,11],ymm0[12],ymm7[13],ymm0[14],ymm7[15]
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
-; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm11, %xmm11
-; AVX512F-SLOW-NEXT: vpbroadcastq 40(%rdi), %xmm14
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm11 = xmm11[0],xmm14[1],xmm11[2,3],xmm14[4],xmm11[5],xmm14[6],xmm11[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm11, %zmm11
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm6
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm6, %ymm14
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm4[3,0,3,0,7,4,7,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm14[1],ymm8[2],ymm14[3],ymm8[4,5],ymm14[6],ymm8[7,8],ymm14[9],ymm8[10],ymm14[11],ymm8[12,13],ymm14[14],ymm8[15]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm10, %xmm10
+; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm8, %xmm8
+; AVX512F-SLOW-NEXT: vpbroadcastq 40(%rdi), %xmm11
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm11[1],xmm8[2,3],xmm11[4],xmm8[5],xmm11[6],xmm8[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm11
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm0, %ymm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm23
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm5[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm8[0],ymm2[1],ymm8[2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8],ymm2[9],ymm8[10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm8 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
+; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm10, %xmm10
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[1,2,2,2]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2],xmm9[3],xmm10[4,5],xmm9[6],xmm10[7]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,0]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm8
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm9, %zmm2
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm2
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm9
; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm10
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm10, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm11, %ymm23
+; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm10, %ymm4
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,1,1]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm19 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpandnq %ymm10, %ymm19, %ymm10
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm10
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm22, %xmm11
-; AVX512F-SLOW-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm12, %xmm12
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0],xmm0[1],xmm12[2],xmm0[3],xmm12[4,5],xmm0[6],xmm12[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,0]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-SLOW-NEXT: vpandnq %ymm10, %ymm21, %ymm10
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm10, %zmm10
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm10
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm12[0],xmm2[0],xmm12[1],xmm2[1],xmm12[2],xmm2[2],xmm12[3],xmm2[3]
+; AVX512F-SLOW-NEXT: vpshufb %xmm13, %xmm4, %xmm4
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,2,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm8[0],xmm2[1],xmm8[2],xmm2[3],xmm8[4,5],xmm2[6],xmm8[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm14, %xmm4
+; AVX512F-SLOW-NEXT: vpbroadcastq 8(%rdi), %xmm7
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm7[1],xmm4[2,3],xmm7[4],xmm4[5],xmm7[6],xmm4[7]
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,2,1,3]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,5,6]
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm7, %zmm4
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm7, %zmm4
+; AVX512F-SLOW-NEXT: vpbroadcastq (%r8), %ymm2
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm9[0,1,1,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm18[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm3, %ymm8
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0,1],ymm4[2],ymm8[3],ymm4[4],ymm8[5,6],ymm4[7],ymm8[8,9],ymm4[10],ymm8[11],ymm4[12],ymm8[13,14],ymm4[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm18[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,2,6,7,6,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm8[1],ymm3[2],ymm8[3],ymm3[4,5],ymm8[6],ymm3[7,8],ymm8[9],ymm3[10],ymm8[11],ymm3[12,13],ymm8[14],ymm3[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,2]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm4
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
+; AVX512F-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm0, %ymm8
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm12[1],ymm8[2,3],ymm12[4],ymm8[5],ymm12[6],ymm8[7,8],ymm12[9],ymm8[10,11],ymm12[12],ymm8[13],ymm12[14],ymm8[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2],ymm5[3,4],ymm0[5,6,7,8],ymm5[9],ymm0[10],ymm5[11,12],ymm0[13,14,15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,2]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm15, %xmm7
-; AVX512F-SLOW-NEXT: vpbroadcastq 8(%rdi), %xmm8
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0],xmm8[1],xmm7[2,3],xmm8[4],xmm7[5],xmm8[6],xmm7[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm24, %xmm8
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm15[0],xmm8[1],xmm15[1],xmm8[2],xmm15[2],xmm8[3],xmm15[3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,2,1,3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,5,6]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm7, %zmm8, %zmm7
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm7
-; AVX512F-SLOW-NEXT: vpbroadcastq (%r8), %ymm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm9[0,1,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm17[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm13, %ymm12
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm12[0,1],ymm7[2],ymm12[3],ymm7[4],ymm12[5,6],ymm7[7],ymm12[8,9],ymm7[10],ymm12[11],ymm7[12],ymm12[13,14],ymm7[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm17[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm13[0],ymm12[1],ymm13[2],ymm12[3],ymm13[4,5],ymm12[6],ymm13[7,8],ymm12[9],ymm13[10],ymm12[11],ymm13[12,13],ymm12[14],ymm13[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm12
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
-; AVX512F-SLOW-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm6, %ymm13
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm4[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3],ymm14[4],ymm13[5],ymm14[6],ymm13[7,8],ymm14[9],ymm13[10,11],ymm14[12],ymm13[13],ymm14[14],ymm13[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[3,2,3,3,7,6,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2],ymm4[3,4],ymm6[5,6,7,8],ymm4[9],ymm6[10],ymm4[11,12],ymm6[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm13, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm12, %zmm8, %zmm4
-; AVX512F-SLOW-NEXT: vpbroadcastq 48(%r8), %ymm6
-; AVX512F-SLOW-NEXT: vpbroadcastq 56(%r8), %ymm8
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm6
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm20, %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm5, %ymm4
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0],ymm4[1],ymm8[2,3],ymm4[4],ymm8[5],ymm4[6],ymm8[7,8],ymm4[9],ymm8[10,11],ymm4[12],ymm8[13],ymm4[14],ymm8[15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm5, %ymm5
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8,9],ymm3[10],ymm5[11],ymm3[12],ymm5[13,14],ymm3[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm21, %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm1, %ymm4
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[3,0,3,0,7,4,7,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8],ymm4[9],ymm5[10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
-; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm1, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm18[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[2,3],ymm5[4],ymm1[5],ymm5[6],ymm1[7,8],ymm5[9],ymm1[10,11],ymm5[12],ymm1[13],ymm5[14],ymm1[15]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm4, %zmm7, %zmm0
+; AVX512F-SLOW-NEXT: vpbroadcastq 48(%r8), %ymm4
+; AVX512F-SLOW-NEXT: vpbroadcastq 56(%r8), %ymm5
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm6, %ymm0
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm20[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5],ymm0[6],ymm5[7,8],ymm0[9],ymm5[10,11],ymm0[12],ymm5[13],ymm0[14],ymm5[15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm20[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm6, %ymm6
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3],ymm5[4],ymm6[5,6],ymm5[7],ymm6[8,9],ymm5[10],ymm6[11],ymm5[12],ymm6[13,14],ymm5[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm23, %ymm5
+; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm1, %ymm5
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm17[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4,5],ymm5[6],ymm6[7,8],ymm5[9],ymm6[10],ymm5[11],ymm6[12,13],ymm5[14],ymm6[15]
+; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3],ymm3[4],ymm1[5],ymm3[6],ymm1[7,8],ymm3[9],ymm1[10,11],ymm3[12],ymm1[13],ymm3[14],ymm1[15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm1
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm3, %zmm16, %zmm1
-; AVX512F-SLOW-NEXT: vpbroadcastq 16(%r8), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpandn %ymm3, %ymm4, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm23, %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm9, %ymm4
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm3, 64(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm6, 256(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, (%r9)
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm1
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm16, %zmm1
+; AVX512F-SLOW-NEXT: vpbroadcastq 16(%r8), %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-SLOW-NEXT: vpandn %ymm0, %ymm3, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm9, %ymm3
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, 64(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, 256(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, (%r9)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm10, 192(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm19, 128(%r9)
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
; AVX512F-FAST-LABEL: store_i16_stride5_vf32:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm11
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm12
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm12[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm6, %xmm3
-; AVX512F-FAST-NEXT: vpbroadcastq 40(%rdi), %xmm4
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5],xmm4[6],xmm3[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm16
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm5, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 32(%rdi), %ymm19
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm19[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm10
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm2, %xmm21
+; AVX512F-FAST-NEXT: vpbroadcastq 40(%rdi), %xmm2
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3],xmm2[4],xmm1[5],xmm2[6],xmm1[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm8
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm13, %ymm2
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm8, %ymm1
; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm20
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7,8],ymm2[9],ymm3[10],ymm2[11],ymm3[12,13],ymm2[14],ymm3[15]
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm8
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm8, %xmm4
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm5 = xmm10[1,2,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2],xmm5[3],xmm4[4,5],xmm5[6],xmm4[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,0]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm19
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm5, %ymm22
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,1,1]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpandn %ymm2, %ymm7, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm18
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm3
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,2,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm9 = xmm3[0],xmm2[1],xmm3[2],xmm2[3],xmm3[4,5],xmm2[6],xmm3[7]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm1, %xmm0
-; AVX512F-FAST-NEXT: vpbroadcastq 8(%rdi), %xmm2
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,4,5,4,5,6,7,12,13]
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm17
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm12[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm11, %ymm1
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6],ymm0[7],ymm1[8,9],ymm0[10],ymm1[11],ymm0[12],ymm1[13,14],ymm0[15]
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29,30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29]
+; AVX512F-FAST-NEXT: vmovdqa64 (%rdx), %ymm18
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm9
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm9[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2],ymm1[3],ymm4[4,5],ymm1[6],ymm4[7,8],ymm1[9],ymm4[10],ymm1[11],ymm4[12,13],ymm1[14],ymm4[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
+; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm4, %xmm13
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm14 = xmm6[1,2,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0],xmm14[1],xmm13[2],xmm14[3],xmm13[4,5],xmm14[6],xmm13[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,0]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm13, %zmm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm14
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm14, %ymm3
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm24
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,1,1]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpandnq %ymm14, %ymm16, %ymm14
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm14, %zmm17
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm17
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm1, %xmm3
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[1,2,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm15[1],xmm3[2],xmm15[3],xmm3[4,5],xmm15[6],xmm3[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1],xmm1[2],xmm14[2],xmm1[3],xmm14[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm7, %xmm22
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm14 = [0,1,0,1,8,9,8,8]
+; AVX512F-FAST-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm1
+; AVX512F-FAST-NEXT: vpbroadcastq 8(%rdi), %xmm3
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3],xmm3[4],xmm1[5],xmm3[6],xmm1[7]
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm10[0],xmm3[1],xmm10[1],xmm3[2],xmm10[2],xmm3[3],xmm10[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,4,5,4,5,6,7,12,13]
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm3, %xmm3
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm7, %xmm23
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm14, %zmm16, %zmm1
+; AVX512F-FAST-NEXT: vpbroadcastq (%r8), %ymm3
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm0[0,1,1,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm10
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm10
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm15
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm15, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm14
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm14[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5],ymm1[6],ymm3[7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13],ymm1[14],ymm3[15]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm15, %ymm13
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm13[0,1],ymm3[2],ymm13[3],ymm3[4],ymm13[5,6],ymm3[7],ymm13[8,9],ymm3[10],ymm13[11],ymm3[12],ymm13[13,14],ymm3[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm3
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm18[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0],ymm1[1],ymm13[2],ymm1[3],ymm13[4,5],ymm1[6],ymm13[7,8],ymm1[9],ymm13[10],ymm1[11],ymm13[12,13],ymm1[14],ymm13[15]
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8],ymm2[9],ymm0[10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
-; AVX512F-FAST-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm14[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5],ymm2[6],ymm0[7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13],ymm2[14],ymm0[15]
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm2 = [26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm12
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm14 = ymm12[0],ymm13[1],ymm12[2],ymm13[3,4],ymm12[5,6,7,8],ymm13[9],ymm12[10],ymm13[11,12],ymm12[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm12
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm13
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm2
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm12[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2],ymm5[3,4],ymm2[5,6,7,8],ymm5[9],ymm2[10],ymm5[11,12],ymm2[13,14,15]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm15, %xmm15
-; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm5, %xmm10
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm8
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm2, %ymm7
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm18[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm11[1],ymm7[2,3],ymm11[4],ymm7[5],ymm11[6],ymm7[7,8],ymm11[9],ymm7[10,11],ymm11[12],ymm7[13],ymm11[14],ymm7[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm20, %zmm7
+; AVX512F-FAST-NEXT: vpbroadcastq 16(%r8), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm3, %ymm13, %ymm3
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm24, %ymm11
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm0
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29,30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29]
+; AVX512F-FAST-NEXT: # ymm13 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm13, %ymm5, %ymm3
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm19[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm7[1],ymm3[2],ymm7[3],ymm3[4,5],ymm7[6],ymm3[7,8],ymm7[9],ymm3[10],ymm7[11],ymm3[12,13],ymm7[14],ymm3[15]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm19[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm5, %ymm5
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1],ymm7[2],ymm5[3],ymm7[4],ymm5[5,6],ymm7[7],ymm5[8,9],ymm7[10],ymm5[11],ymm7[12],ymm5[13,14],ymm7[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,3,2,3,10,11,10,10]
+; AVX512F-FAST-NEXT: vpermi2q %zmm3, %zmm5, %zmm7
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm3 = [26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm8, %ymm5
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm9[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm11[1],ymm5[2],ymm11[3,4],ymm5[5,6,7,8],ymm11[9],ymm5[10],ymm11[11,12],ymm5[13,14,15]
; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm8, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm5[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2],ymm11[3],ymm1[4,5],ymm11[6],ymm1[7,8],ymm11[9],ymm1[10],ymm11[11],ymm1[12,13],ymm11[14],ymm1[15]
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm11
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm6, %xmm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm21, %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm8, %ymm6
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm5[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2,3],ymm6[4],ymm11[5],ymm6[6],ymm11[7,8],ymm6[9],ymm11[10,11],ymm6[12],ymm11[13],ymm6[14],ymm11[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm15[0,1,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,0]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm8, %ymm8
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm8[0,1],ymm5[2],ymm8[3],ymm5[4],ymm8[5,6],ymm5[7],ymm8[8,9],ymm5[10],ymm8[11],ymm5[12],ymm8[13,14],ymm5[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm20, %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm13, %ymm6
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm12[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4,5],ymm6[6],ymm8[7,8],ymm6[9],ymm8[10],ymm6[11],ymm8[12,13],ymm6[14],ymm8[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm23, %ymm8
-; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm13, %ymm8
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm12[1],ymm8[2,3],ymm12[4],ymm8[5],ymm12[6],ymm8[7,8],ymm12[9],ymm8[10,11],ymm12[12],ymm8[13],ymm12[14],ymm8[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm8
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm8[0,1,1,1]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm22, %ymm13
-; AVX512F-FAST-NEXT: vpshufb %ymm13, %ymm8, %ymm8
-; AVX512F-FAST-NEXT: vpbroadcastq 16(%r8), %ymm13
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandn %ymm13, %ymm15, %ymm13
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm19
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm18
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm11, %zmm9
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm11 = zmm17[0,1,0,1,4,5,4,5]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm9, %zmm13, %zmm11
-; AVX512F-FAST-NEXT: vpbroadcastq (%r8), %ymm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm9
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm3
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm13, %zmm0
-; AVX512F-FAST-NEXT: vpbroadcastq 48(%r8), %ymm3
-; AVX512F-FAST-NEXT: vpbroadcastq 56(%r8), %ymm7
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm3
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512F-FAST-NEXT: vpbroadcastq 24(%r8), %ymm0
-; AVX512F-FAST-NEXT: vpbroadcastq 32(%r8), %ymm4
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm2, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 64(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 128(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, 256(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, (%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm18, 192(%r9)
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm9[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm8[1],ymm1[2,3],ymm8[4],ymm1[5],ymm8[6],ymm1[7,8],ymm8[9],ymm1[10,11],ymm8[12],ymm1[13],ymm8[14],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [2,3,2,3,10,10,11,10]
+; AVX512F-FAST-NEXT: vpermi2q %zmm5, %zmm1, %zmm8
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm7, %zmm16, %zmm8
+; AVX512F-FAST-NEXT: vpbroadcastq 48(%r8), %ymm1
+; AVX512F-FAST-NEXT: vpbroadcastq 56(%r8), %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm22, %xmm5
+; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm4, %xmm4
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3,4],ymm2[5,6,7,8],ymm3[9],ymm2[10],ymm3[11,12],ymm2[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [2,2,3,2,8,9,8,9]
+; AVX512F-FAST-NEXT: vpermi2q %zmm4, %zmm2, %zmm3
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm21, %xmm2
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm12[0],xmm2[0],xmm12[1],xmm2[1],xmm12[2],xmm2[2],xmm12[3],xmm2[3]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm23, %xmm4
+; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm2
+; AVX512F-FAST-NEXT: vpshufb %ymm13, %ymm15, %ymm4
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm14[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7,8],ymm5[9],ymm4[10],ymm5[11],ymm4[12,13],ymm5[14],ymm4[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [2,3,2,2,8,9,8,9]
+; AVX512F-FAST-NEXT: vpermi2q %zmm2, %zmm4, %zmm5
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm20, %zmm5
+; AVX512F-FAST-NEXT: vpbroadcastq 24(%r8), %ymm2
+; AVX512F-FAST-NEXT: vpbroadcastq 32(%r8), %ymm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm2
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 256(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 64(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, (%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm17, 192(%r9)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
@@ -5698,675 +5688,633 @@ define void @store_i16_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-SLOW-LABEL: store_i16_stride5_vf64:
; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: subq $616, %rsp # imm = 0x268
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rcx), %ymm13
+; AVX512F-SLOW-NEXT: subq $520, %rsp # imm = 0x208
+; AVX512F-SLOW-NEXT: vmovdqa 96(%rcx), %ymm15
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa %ymm1, %ymm10
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdx), %ymm6
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm15, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa %ymm1, %ymm8
+; AVX512F-SLOW-NEXT: vmovdqa64 96(%rdx), %ymm16
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[3,0,3,0,7,4,7,4]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
; AVX512F-SLOW-NEXT: vmovdqa 96(%rcx), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm0, %xmm3
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdx), %xmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[1,2,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2],xmm4[3],xmm3[4,5],xmm4[6],xmm3[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,0]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm0, %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa 96(%rdx), %xmm3
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,2,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2],xmm4[3],xmm2[4,5],xmm4[6],xmm2[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rsi), %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
-; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm7, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm16
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdi), %ymm8
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm8[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5],ymm1[6],ymm3[7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13],ymm1[14],ymm3[15]
+; AVX512F-SLOW-NEXT: vmovdqa 96(%rsi), %ymm2
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm2, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa64 96(%rdi), %ymm18
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm18[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm1[1],ymm4[2,3],ymm1[4],ymm4[5],ymm1[6],ymm4[7,8],ymm1[9],ymm4[10,11],ymm1[12],ymm4[13],ymm1[14],ymm4[15]
; AVX512F-SLOW-NEXT: vmovdqa 96(%rsi), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm1, %xmm4
-; AVX512F-SLOW-NEXT: vpbroadcastq 104(%rdi), %xmm9
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm9[1],xmm4[2,3],xmm9[4],xmm4[5],xmm9[6],xmm4[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 96(%r8), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm3, %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm5, %ymm22
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,1,1]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpandn %ymm3, %ymm5, %ymm3
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 (%rdx), %ymm17
-; AVX512F-SLOW-NEXT: vmovdqa64 64(%rdx), %ymm24
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm24[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm1, %xmm5
+; AVX512F-SLOW-NEXT: vpbroadcastq 104(%rdi), %xmm6
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3],xmm6[4],xmm5[5],xmm6[6],xmm5[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 96(%r8), %ymm4
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
+; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm4, %ymm5
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm6, %ymm29
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,1,1]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-SLOW-NEXT: vpandnq %ymm4, %ymm17, %ymm4
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm0, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm13
+; AVX512F-SLOW-NEXT: vmovdqa 64(%rdx), %ymm3
+; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,2,3,3,7,6,7,7]
; AVX512F-SLOW-NEXT: vmovdqa 64(%rcx), %ymm4
; AVX512F-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3,4],ymm4[5,6,7,8],ymm3[9],ymm4[10],ymm3[11,12],ymm4[13,14,15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,2]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm20
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm20[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8],ymm0[9],ymm2[10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,2]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm0[0,1,0,1]
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa 96(%rdi), %xmm0
; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,5,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa64 64(%rdi), %ymm19
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm19[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %ymm3
+; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,2,6,7,6,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4,5],ymm1[6],ymm3[7,8],ymm1[9],ymm3[10],ymm1[11],ymm3[12,13],ymm1[14],ymm3[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,2]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm1[0,1,2,3],zmm0[0,1,0,1]
; AVX512F-SLOW-NEXT: vmovdqa 64(%rdx), %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm7
; AVX512F-SLOW-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm23
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm1, %xmm1
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm1, %xmm1
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,2,2]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4,5],xmm0[6],xmm1[7]
; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %xmm1
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm1, %xmm2
-; AVX512F-SLOW-NEXT: vpbroadcastq 72(%rdi), %xmm4
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3],xmm4[4],xmm2[5],xmm4[6],xmm2[7]
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rdi), %xmm4
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,5,6]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,2,3,3,7,6,7,7]
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm11
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm11[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3,4],ymm2[5,6,7,8],ymm1[9],ymm2[10],ymm1[11,12],ymm2[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,2]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm0, %xmm1
-; AVX512F-SLOW-NEXT: vpbroadcastq 8(%rdi), %xmm2
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3],xmm2[4],xmm1[5],xmm2[6],xmm1[7]
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %xmm0
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm0, %xmm1
+; AVX512F-SLOW-NEXT: vpbroadcastq 72(%rdi), %xmm3
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3],xmm3[4],xmm1[5],xmm3[6],xmm1[7]
+; AVX512F-SLOW-NEXT: vmovdqa 64(%rdi), %xmm3
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm4
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,5,6]
; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm0, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm1, %xmm1
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm1[0,1,0,1]
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm3[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vmovdqa64 (%rdi), %ymm21
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm21[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm9[1],ymm2[2],ymm9[3],ymm2[4,5],ymm9[6],ymm2[7,8],ymm9[9],ymm2[10],ymm9[11],ymm2[12,13],ymm9[14],ymm2[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm19
-; AVX512F-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm18
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512F-SLOW-NEXT: vpshufb %ymm10, %ymm2, %ymm9
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm18[3,0,3,0,7,4,7,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2],ymm9[3],ymm10[4,5],ymm9[6],ymm10[7,8],ymm9[9],ymm10[10],ymm9[11],ymm10[12,13],ymm9[14],ymm10[15]
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm4, %xmm4
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm4[0],xmm0[1],xmm4[2],xmm0[3],xmm4[4,5],xmm0[6],xmm4[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,0]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm4, %xmm0
+; AVX512F-SLOW-NEXT: vpbroadcastq 8(%rdi), %xmm1
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]
+; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,5,6]
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm16, %ymm0
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm9
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm4[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5],ymm9[6],ymm10[7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13],ymm9[14],ymm10[15]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm9
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vpbroadcastq 40(%rdi), %xmm14
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3],xmm14[4],xmm0[5],xmm14[6],xmm0[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm29
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm0[0,1,1,1]
-; AVX512F-SLOW-NEXT: vpandn %ymm10, %ymm5, %ymm5
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm22, %ymm10
-; AVX512F-SLOW-NEXT: vpshufb %ymm10, %ymm0, %ymm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm25
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm23, %xmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm27
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm5
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
-; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm14, %xmm12
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm12, %ymm30
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,2,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm5[1],xmm0[2],xmm5[3],xmm0[4,5],xmm5[6],xmm0[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm8[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm7, %ymm5
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0,1],ymm0[2],ymm5[3],ymm0[4],ymm5[5,6],ymm0[7],ymm5[8,9],ymm0[10],ymm5[11],ymm0[12],ymm5[13,14],ymm0[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm26
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm7[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm0[0],ymm5[1],ymm0[2],ymm5[3],ymm0[4,5],ymm5[6],ymm0[7,8],ymm5[9],ymm0[10],ymm5[11],ymm0[12,13],ymm5[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
-; AVX512F-SLOW-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm13, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm6[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm0[0],ymm7[1],ymm0[2,3],ymm7[4],ymm0[5],ymm7[6],ymm0[7,8],ymm7[9],ymm0[10,11],ymm7[12],ymm0[13],ymm7[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm13[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm6[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[3,2,3,3,7,6,7,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm0[0],ymm6[1],ymm0[2],ymm6[3,4],ymm0[5,6,7,8],ymm6[9],ymm0[10],ymm6[11,12],ymm0[13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm1, %ymm6
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm6[0,1],ymm0[2],ymm6[3],ymm0[4],ymm6[5,6],ymm0[7],ymm6[8,9],ymm0[10],ymm6[11],ymm0[12],ymm6[13,14],ymm0[15]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3,4],ymm0[5,6,7,8],ymm3[9],ymm0[10],ymm3[11,12],ymm0[13,14,15]
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,2]
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm1, %xmm1
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa64 (%rdi), %ymm30
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm9[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,2,6,7,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm2, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13],ymm1[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm30[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,2]
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,5,6]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm0[0,1,2,3],zmm1[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm22
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm1
+; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm22[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2],ymm0[3],ymm8[4,5],ymm0[6],ymm8[7,8],ymm0[9],ymm8[10],ymm0[11],ymm8[12,13],ymm0[14],ymm8[15]
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm7, %xmm7
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,2,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0],xmm3[1],xmm7[2],xmm3[3],xmm7[4,5],xmm3[6],xmm7[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,0]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm0, %ymm7
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2,3],ymm7[4],ymm8[5],ymm7[6],ymm8[7,8],ymm7[9],ymm8[10,11],ymm7[12],ymm8[13],ymm7[14],ymm8[15]
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm4, %xmm4
+; AVX512F-SLOW-NEXT: vpbroadcastq 40(%rdi), %xmm8
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm8[1],xmm4[2,3],xmm8[4],xmm4[5],xmm8[6],xmm4[7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm25
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm4
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm4[0,1,1,1]
+; AVX512F-SLOW-NEXT: vpandnq %ymm7, %ymm17, %ymm7
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm29, %ymm8
+; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm4, %ymm4
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm17
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm11, %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm7
+; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3]
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm8, %xmm5
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm5, %ymm27
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm4, %xmm4
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm7[1,2,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2],xmm5[3],xmm4[4,5],xmm5[6],xmm4[7]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm26
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm18[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm2, %ymm5
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1],ymm4[2],ymm5[3],ymm4[4],ymm5[5,6],ymm4[7],ymm5[8,9],ymm4[10],ymm5[11],ymm4[12],ymm5[13,14],ymm4[15]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm21
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,3,2,2,6,7,6,6]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm18[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm11 = ymm2[0],ymm4[1],ymm2[2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7,8],ymm4[9],ymm2[10],ymm4[11],ymm2[12,13],ymm4[14],ymm2[15]
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
+; AVX512F-SLOW-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm15, %ymm4
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm16[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm4[0],ymm5[1],ymm4[2,3],ymm5[4],ymm4[5],ymm5[6],ymm4[7,8],ymm5[9],ymm4[10,11],ymm5[12],ymm4[13],ymm5[14],ymm4[15]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm16[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm15 = ymm4[0],ymm5[1],ymm4[2],ymm5[3,4],ymm4[5,6,7,8],ymm5[9],ymm4[10],ymm5[11,12],ymm4[13,14,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm3[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm0, %ymm5
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm4[2],ymm5[3],ymm4[4],ymm5[5,6],ymm4[7],ymm5[8,9],ymm4[10],ymm5[11],ymm4[12],ymm5[13,14],ymm4[15]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,5,7,8,9,10,11,15,14,13,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,2,6,7,6,6]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm5 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4,5],ymm3[6],ymm0[7,8],ymm3[9],ymm0[10],ymm3[11],ymm0[12,13],ymm3[14],ymm0[15]
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm22[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5],ymm3[6],ymm0[7,8],ymm3[9],ymm0[10,11],ymm3[12],ymm0[13],ymm3[14],ymm0[15]
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[3,2,3,3,7,6,7,7]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm22[3,2,3,3,7,6,7,7]
; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm10 = ymm0[0],ymm1[1],ymm0[2],ymm1[3,4],ymm0[5,6,7,8],ymm1[9],ymm0[10],ymm1[11,12],ymm0[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm16, %ymm13
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm21[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm3, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm21[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3],ymm3[4],ymm1[5,6],ymm3[7],ymm1[8,9],ymm3[10],ymm1[11],ymm3[12],ymm1[13,14],ymm3[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm18
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm11, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,0,3,0,7,4,7,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm11, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3],ymm3[4],ymm1[5],ymm3[6],ymm1[7,8],ymm3[9],ymm1[10,11],ymm3[12],ymm1[13],ymm3[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm11
-; AVX512F-SLOW-NEXT: vpbroadcastq 16(%r8), %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpandnq %ymm0, %ymm21, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm22, %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm0, %ymm3
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm16
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm15, %ymm1
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm20[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5],ymm1[6],ymm3[7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13],ymm1[14],ymm3[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm19[2,3,2,2]
-; AVX512F-SLOW-NEXT: vprolq $16, %ymm15, %ymm13
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm20[0,1,2,1,4,5,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3],ymm15[4],ymm13[5,6],ymm15[7],ymm13[8,9],ymm15[10],ymm13[11],ymm15[12],ymm13[13,14],ymm15[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm1, %zmm20
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm13, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm24[3,0,3,0,7,4,7,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm15 = ymm15[0],ymm1[1],ymm15[2],ymm1[3],ymm15[4,5],ymm1[6],ymm15[7,8],ymm1[9],ymm15[10],ymm1[11],ymm15[12,13],ymm1[14],ymm15[15]
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm13, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm9[0,2,1,3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,4,5,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm24[1,1,1,2,5,5,5,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0],ymm9[1],ymm1[2,3],ymm9[4],ymm1[5],ymm9[6],ymm1[7,8],ymm9[9],ymm1[10,11],ymm9[12],ymm1[13],ymm9[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, (%rsp), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm17 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm24, %zmm17
+; AVX512F-SLOW-NEXT: vmovdqa %ymm12, %ymm2
+; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm9, %ymm1
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm30[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm14[0],ymm1[1],ymm14[2,3],ymm1[4],ymm14[5],ymm1[6],ymm14[7,8],ymm1[9],ymm14[10,11],ymm1[12],ymm14[13],ymm1[14],ymm14[15]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm9, %ymm9
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm30[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm14[2],ymm9[3],ymm14[4],ymm9[5,6],ymm14[7],ymm9[8,9],ymm14[10],ymm9[11],ymm14[12],ymm9[13,14],ymm14[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm23
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm6, %ymm1
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm13[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm9[0],ymm1[1],ymm9[2],ymm1[3],ymm9[4,5],ymm1[6],ymm9[7,8],ymm1[9],ymm9[10],ymm1[11],ymm9[12,13],ymm1[14],ymm9[15]
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm6, %ymm6
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm13[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm9[1],ymm6[2,3],ymm9[4],ymm6[5],ymm9[6],ymm6[7,8],ymm9[9],ymm6[10,11],ymm9[12],ymm6[13],ymm9[14],ymm6[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm6
+; AVX512F-SLOW-NEXT: vpbroadcastq 16(%r8), %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm18 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-SLOW-NEXT: vpandnq %ymm1, %ymm18, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm9
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm29, %ymm3
+; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm9, %ymm13
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm13, %zmm13
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm12, %ymm1
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm19[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm14[0],ymm1[1],ymm14[2,3],ymm1[4],ymm14[5],ymm1[6],ymm14[7,8],ymm1[9],ymm14[10,11],ymm1[12],ymm14[13],ymm1[14],ymm14[15]
+; AVX512F-SLOW-NEXT: vprolq $16, %ymm12, %ymm12
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm19[0,1,2,1,4,5,6,5]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3],ymm14[4],ymm12[5,6],ymm14[7],ymm12[8,9],ymm14[10],ymm12[11],ymm14[12],ymm12[13,14],ymm14[15]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm1, %zmm30
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm12[3,0,3,0,7,4,7,4]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm1[1],ymm14[2],ymm1[3],ymm14[4,5],ymm1[6],ymm14[7,8],ymm1[9],ymm14[10],ymm1[11],ymm14[12,13],ymm1[14],ymm14[15]
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm2, %ymm1
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[1,1,1,2,5,5,5,6]
+; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5],ymm2[6],ymm1[7,8],ymm2[9],ymm1[10,11],ymm2[12],ymm1[13],ymm2[14],ymm1[15]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm31 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: vpbroadcastq 88(%r8), %ymm1
-; AVX512F-SLOW-NEXT: vpbroadcastq 96(%r8), %ymm22
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm22 = [0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm17, %zmm22, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm31, %zmm2, %zmm17
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm5
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm17, %zmm24, %zmm5
-; AVX512F-SLOW-NEXT: vpbroadcastq 24(%r8), %ymm3
-; AVX512F-SLOW-NEXT: vpbroadcastq 32(%r8), %ymm17
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm3, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm22, %zmm3
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm27[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpermq $4, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm17 = mem[0,1,0,0]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm13
-; AVX512F-SLOW-NEXT: vpbroadcastq 80(%r8), %ymm15
-; AVX512F-SLOW-NEXT: vpandnq %ymm15, %ymm21, %ymm15
-; AVX512F-SLOW-NEXT: vmovdqa 64(%r8), %ymm9
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm9, %ymm5
+; AVX512F-SLOW-NEXT: vpbroadcastq 96(%r8), %ymm14
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm29
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm19 = [0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm31, %zmm19, %zmm29
+; AVX512F-SLOW-NEXT: vpternlogq $226, (%rsp), %zmm16, %zmm28 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: vpbroadcastq 24(%r8), %ymm14
+; AVX512F-SLOW-NEXT: vpbroadcastq 32(%r8), %ymm22
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm14, %zmm14
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm28, %zmm19, %zmm14
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm24[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpermq $4, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm22 = mem[0,1,0,0]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vpbroadcastq 80(%r8), %ymm2
+; AVX512F-SLOW-NEXT: vpandnq %ymm2, %ymm18, %ymm2
+; AVX512F-SLOW-NEXT: vmovdqa 64(%r8), %ymm12
+; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm12, %ymm1
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,1,1]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm27[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[0,1,0,0]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm30[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm28[0,1,0,0]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm26[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm23[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm10[2,2,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm5, %zmm5
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm21[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,2]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,2]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm20[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,2]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm10[2,2,3,2]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0]
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm21 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm25 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm10 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm29 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm15, %zmm10, %zmm23
-; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm15, %zmm29, %zmm25
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm19, %zmm15
-; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: # zmm17 = mem[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm15, %zmm19, %zmm17
-; AVX512F-SLOW-NEXT: vpbroadcastq 64(%r8), %ymm15
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm15, %zmm9
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm17, %zmm15, %zmm9
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm17
-; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: # zmm21 = mem[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm17, %zmm19, %zmm21
-; AVX512F-SLOW-NEXT: vpbroadcastq (%r8), %ymm17
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm17, %zmm0
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm15, %zmm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm27, %zmm14
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm12, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm14, %zmm19, %zmm8
-; AVX512F-SLOW-NEXT: vpbroadcastq 112(%r8), %ymm12
-; AVX512F-SLOW-NEXT: vpbroadcastq 120(%r8), %ymm14
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm12
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm8, %zmm14, %zmm12
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm19, %zmm2
+; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm2, %zmm21, %zmm10
+; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm2, %zmm25, %zmm17
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm19, %zmm2
+; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm19 = mem[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm21, %zmm19
+; AVX512F-SLOW-NEXT: vpbroadcastq 64(%r8), %ymm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm2, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm19, %zmm12, %zmm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm18, %zmm18
+; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm19 = mem[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm18, %zmm21, %zmm19
+; AVX512F-SLOW-NEXT: vpbroadcastq (%r8), %ymm18
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm18, %zmm9
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm19, %zmm12, %zmm9
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm28, %zmm11
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm8, %zmm8
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm11, %zmm21, %zmm8
+; AVX512F-SLOW-NEXT: vpbroadcastq 112(%r8), %ymm11
+; AVX512F-SLOW-NEXT: vpbroadcastq 120(%r8), %ymm12
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0]
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm8, %zmm12, %zmm11
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm21, %zmm3
; AVX512F-SLOW-NEXT: vpbroadcastq 48(%r8), %ymm4
-; AVX512F-SLOW-NEXT: vpbroadcastq 56(%r8), %ymm6
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm14, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm18, %zmm24, %zmm11
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm20, %zmm24, %zmm13
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm2, %zmm11, %zmm16
-; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm2, %zmm13, %zmm5
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm5, 384(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm16, 64(%r9)
+; AVX512F-SLOW-NEXT: vpbroadcastq 56(%r8), %ymm5
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm12, %zmm4
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm23, %zmm16, %zmm6
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm30, %zmm16, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm3, %zmm6, %zmm13
+; AVX512F-SLOW-NEXT: vpternlogq $248, %zmm3, %zmm0, %zmm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm1, 384(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm13, 64(%r9)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, 256(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm12, 576(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, (%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm25, 192(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm3, 128(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm9, 320(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm1, 448(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm23, 512(%r9)
-; AVX512F-SLOW-NEXT: addq $616, %rsp # imm = 0x268
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm11, 576(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm9, (%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm17, 192(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm14, 128(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, 320(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm29, 448(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm10, 512(%r9)
+; AVX512F-SLOW-NEXT: addq $520, %rsp # imm = 0x208
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
; AVX512F-FAST-LABEL: store_i16_stride5_vf64:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: subq $456, %rsp # imm = 0x1C8
-; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %ymm8
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm8, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa %ymm1, %ymm9
-; AVX512F-FAST-NEXT: vmovdqa64 96(%rdx), %ymm20
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %xmm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm2
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm3, %xmm31
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm13
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm13[1,2,2,2]
+; AVX512F-FAST-NEXT: subq $360, %rsp # imm = 0x168
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %ymm6
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 96(%rdx), %ymm18
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512F-FAST-NEXT: vmovdqa 96(%rcx), %xmm2
+; AVX512F-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = <10,11,u,u,6,7,u,u,8,9,8,9,u,u,8,9>
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdx), %xmm3
+; AVX512F-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,2,2,2]
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2],xmm3[3],xmm2[4,5],xmm3[6],xmm2[7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %ymm10
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm10, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa %ymm2, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %ymm12
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm12[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5],ymm0[6],ymm2[7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa64 96(%rdi), %ymm23
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm23[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5],ymm1[6],ymm3[7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13],ymm1[14],ymm3[15]
; AVX512F-FAST-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm4, %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm4, %xmm17
-; AVX512F-FAST-NEXT: vpbroadcastq 104(%rdi), %xmm4
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5],xmm4[6],xmm3[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm26
-; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,1,1]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpandnq %ymm2, %ymm22, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm4
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm7, %xmm16
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = <6,7,u,u,10,11,6,7,u,u,8,9,u,u,12,13>
; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm4
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,2,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2],xmm2[3],xmm4[4,5],xmm2[6],xmm4[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
-; AVX512F-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm2
+; AVX512F-FAST-NEXT: vpbroadcastq 104(%rdi), %xmm5
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3],xmm5[4],xmm4[5],xmm5[6],xmm4[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 96(%r8), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
+; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm3, %ymm4
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm3[0,1,1,1]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm5, %ymm3, %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512F-FAST-NEXT: vmovdqu %ymm4, (%rsp) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm4, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa64 32(%rdx), %ymm27
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm27[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8],ymm4[9],ymm5[10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm5, %xmm5
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,2,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm7[1],xmm5[2],xmm7[3],xmm5[4,5],xmm7[6],xmm5[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,0]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm4, %xmm5
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm5
; AVX512F-FAST-NEXT: vpbroadcastq 72(%rdi), %xmm7
; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm7[1],xmm5[2,3],xmm7[4],xmm5[5],xmm7[6],xmm5[7]
; AVX512F-FAST-NEXT: vmovdqa 64(%rdi), %xmm7
; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,8,9,10,11,4,5,4,5,6,7,12,13]
; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm4, %xmm4
-; AVX512F-FAST-NEXT: vmovdqa %xmm7, %xmm11
+; AVX512F-FAST-NEXT: vmovdqa %xmm7, %xmm10
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm14
-; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm14, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 32(%rdx), %ymm24
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm24[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8],ymm4[9],ymm5[10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm5
-; AVX512F-FAST-NEXT: vpbroadcastq 8(%rdi), %xmm7
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm7[1],xmm5[2,3],xmm7[4],xmm5[5],xmm7[6],xmm5[7]
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
-; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm2, %xmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm9
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm9, %xmm2
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm15
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm15[1,2,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm7[1],xmm2[2],xmm7[3],xmm2[4,5],xmm7[6],xmm2[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,0]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm5, %xmm7
+; AVX512F-FAST-NEXT: vpbroadcastq 8(%rdi), %xmm9
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3],xmm9[4],xmm7[5],xmm9[6],xmm7[7]
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm5, %xmm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm13, %ymm7
; AVX512F-FAST-NEXT: vmovdqa64 32(%rdi), %ymm30
-; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm2, %ymm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm30[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm7[0],ymm4[1],ymm7[2,3],ymm4[4],ymm7[5],ymm4[6],ymm7[7,8],ymm4[9],ymm7[10,11],ymm4[12],ymm7[13],ymm4[14],ymm7[15]
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm11, %xmm0
-; AVX512F-FAST-NEXT: vpbroadcastq 40(%rdi), %xmm7
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm7[1],xmm0[2,3],xmm7[4],xmm0[5],xmm7[6],xmm0[7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm28
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm0
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,1,1,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm0
-; AVX512F-FAST-NEXT: vpandnq %ymm4, %ymm22, %ymm3
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm27
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm5
-; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm4, %xmm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4,5],xmm0[6],xmm1[7]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm12[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm10, %ymm1
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6],ymm0[7],ymm1[8,9],ymm0[10],ymm1[11],ymm0[12],ymm1[13,14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29,30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29]
-; AVX512F-FAST-NEXT: # ymm5 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm10, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm12[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
-; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm8, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa %ymm1, %ymm12
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3],ymm1[4],ymm0[5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10,11],ymm1[12],ymm0[13],ymm1[14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm7 = [26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31]
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm8, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3,4],ymm0[5,6,7,8],ymm1[9],ymm0[10],ymm1[11,12],ymm0[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512F-FAST-NEXT: vmovdqa64 64(%rdx), %ymm20
-; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %ymm6
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm6, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3,4],ymm0[5,6,7,8],ymm1[9],ymm0[10],ymm1[11,12],ymm0[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm31, %xmm0
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm0[0],xmm13[0],xmm0[1],xmm13[1],xmm0[2],xmm13[2],xmm0[3],xmm13[3]
-; AVX512F-FAST-NEXT: vmovdqa64 64(%rdi), %ymm29
-; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm29[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4,5],ymm8[6],ymm0[7,8],ymm8[9],ymm0[10],ymm8[11],ymm0[12,13],ymm8[14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm31
-; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm17, %xmm1
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm30[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm2, %ymm8
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1],ymm0[2],ymm8[3],ymm0[4],ymm8[5,6],ymm0[7],ymm8[8,9],ymm0[10],ymm8[11],ymm0[12],ymm8[13,14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm30[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8],ymm2[9],ymm0[10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm14, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm24[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5],ymm2[6],ymm0[7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13],ymm2[14],ymm0[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm14, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm24[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm14 = ymm0[0],ymm2[1],ymm0[2],ymm2[3,4],ymm0[5,6,7,8],ymm2[9],ymm0[10],ymm2[11,12],ymm0[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa64 (%rdx), %ymm30
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm30[3,2,3,3,7,6,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm7[1],ymm2[2],ymm7[3,4],ymm2[5,6,7,8],ymm7[9],ymm2[10],ymm7[11,12],ymm2[13,14,15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm9[0],xmm15[0],xmm9[1],xmm15[1],xmm9[2],xmm15[2],xmm9[3],xmm15[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm4, %xmm8
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm15
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm0
-; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm9
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm9[2,3,2,3,6,7,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2],ymm5[3],ymm2[4,5],ymm5[6],ymm2[7,8],ymm5[9],ymm2[10],ymm5[11],ymm2[12,13],ymm5[14],ymm2[15]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm17
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,8,9,10,11,4,5,4,5,6,7,12,13]
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm7
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = <u,u,0,1,u,u,u,u,14,15,u,u,2,3,u,u,u,u,16,17,u,u,u,u,30,31,u,u,18,19,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm11
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm9[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm11 = ymm13[0],ymm11[1],ymm13[2,3],ymm11[4],ymm13[5],ymm11[6],ymm13[7,8],ymm11[9],ymm13[10,11],ymm11[12],ymm13[13],ymm11[14],ymm13[15]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm0, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3],ymm9[4],ymm0[5,6],ymm9[7],ymm0[8,9],ymm9[10],ymm0[11],ymm9[12],ymm0[13,14],ymm9[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm24
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,12,13,u,u,0,1,u,u,u,u,14,15,u,u,u,u,28,29,u,u,16,17,u,u,u,u,30,31,u,u>
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm30[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2],ymm0[3],ymm9[4,5],ymm0[6],ymm9[7,8],ymm0[9],ymm9[10],ymm0[11],ymm9[12,13],ymm0[14],ymm9[15]
-; AVX512F-FAST-NEXT: vmovdqa %ymm12, %ymm4
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm1
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm30[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm9[1],ymm1[2,3],ymm9[4],ymm1[5],ymm9[6],ymm1[7,8],ymm9[9],ymm1[10,11],ymm9[12],ymm1[13],ymm9[14],ymm1[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm30[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm9[0],ymm7[1],ymm9[2,3],ymm7[4],ymm9[5],ymm7[6],ymm9[7,8],ymm7[9],ymm9[10,11],ymm7[12],ymm9[13],ymm7[14],ymm9[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512F-FAST-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512F-FAST-NEXT: vpbroadcastq 40(%rdi), %xmm9
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm9[1],xmm1[2,3],xmm9[4],xmm1[5],xmm9[6],xmm1[7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm16
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm1[0,1,1,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vpandn %ymm7, %ymm3, %ymm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm29
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm1, %xmm3
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %xmm7
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm7[1,2,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm9[1],xmm3[2],xmm9[3],xmm3[4,5],xmm9[6],xmm3[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa %xmm7, %xmm9
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,1,0,1,8,9,8,8]
+; AVX512F-FAST-NEXT: vpermt2q %zmm3, %zmm7, %zmm1
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm19 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm31, %zmm19
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,2,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2],xmm4[3],xmm0[4,5],xmm4[6],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm7, %zmm1
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm21 = zmm5[0,1,0,1,4,5,4,5]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm31, %zmm21
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa64 (%rdi), %ymm28
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm3, %ymm0
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm28[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm28[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm3, %ymm3
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3],ymm1[4],ymm3[5,6],ymm1[7],ymm3[8,9],ymm1[10],ymm3[11],ymm1[12],ymm3[13,14],ymm1[15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm10
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm4, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25,18,19,0,0,22,23,22,23,0,0,20,21,0,0,24,25]
+; AVX512F-FAST-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm4, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3],ymm3[4],ymm1[5],ymm3[6],ymm1[7,8],ymm3[9],ymm1[10,11],ymm3[12],ymm1[13],ymm3[14],ymm1[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm25
; AVX512F-FAST-NEXT: vpbroadcastq 16(%r8), %ymm0
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandnq %ymm0, %ymm30, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm11
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,12,13,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm13
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm13, %zmm13
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm0
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm29[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm12[0],ymm0[1],ymm12[2,3],ymm0[4],ymm12[5],ymm0[6],ymm12[7,8],ymm0[9],ymm12[10,11],ymm0[12],ymm12[13],ymm0[14],ymm12[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm16[0,1,0,0]
-; AVX512F-FAST-NEXT: vprolq $16, %ymm3, %ymm3
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm29[0,1,2,1,4,5,6,5]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm10[2],ymm3[3],ymm10[4],ymm3[5,6],ymm10[7],ymm3[8,9],ymm10[10],ymm3[11],ymm10[12],ymm3[13,14],ymm10[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm3
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm20[3,0,3,0,7,4,7,4]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm10[0],ymm3[1],ymm10[2],ymm3[3],ymm10[4,5],ymm3[6],ymm10[7,8],ymm3[9],ymm10[10],ymm3[11],ymm10[12,13],ymm3[14],ymm10[15]
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm11[0,1,1,1]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm20[1,1,1,2,5,5,5,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm11[1],ymm4[2,3],ymm11[4],ymm4[5],ymm11[6],ymm4[7,8],ymm11[9],ymm4[10,11],ymm11[12],ymm4[13],ymm11[14],ymm4[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %ymm4
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm4[0,1,1,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm4
-; AVX512F-FAST-NEXT: vpbroadcastq 80(%r8), %ymm20
-; AVX512F-FAST-NEXT: vpandnq %ymm20, %ymm30, %ymm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm4
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm26 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm28 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vmovdqu64 (%rsp), %zmm29 # 64-byte Reload
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm1, %zmm26, %zmm29
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm1, %zmm28, %zmm27
-; AVX512F-FAST-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm1 = mem[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq $174, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm16 = mem[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: # ymm20 = mem[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm25[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm8[0,1,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm31[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm7[0,1,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm22[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm19[2,3,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm18[2,2,3,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm17[2,3,2,2]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm17 # 32-byte Folded Reload
-; AVX512F-FAST-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm18 = mem[0,1,0,1,4,5,4,5]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm17, %zmm19, %zmm18
-; AVX512F-FAST-NEXT: vpbroadcastq 64(%r8), %ymm17
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm17, %zmm11
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm18, %zmm17, %zmm11
-; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm18, %zmm12
-; AVX512F-FAST-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm18 = mem[0,1,0,1,4,5,4,5]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm12, %zmm19, %zmm18
-; AVX512F-FAST-NEXT: vpbroadcastq (%r8), %ymm12
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm18, %zmm17, %zmm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm1, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm20, %zmm12
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm19, %zmm12
-; AVX512F-FAST-NEXT: vpbroadcastq 112(%r8), %ymm1
-; AVX512F-FAST-NEXT: vpbroadcastq 120(%r8), %ymm16
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm1, %zmm1
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm12, %zmm16, %zmm1
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm26, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm28, %zmm2
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm12, %zmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm21, %zmm6
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm31, %zmm14
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm6, %zmm19, %zmm14
-; AVX512F-FAST-NEXT: vpbroadcastq 48(%r8), %ymm6
-; AVX512F-FAST-NEXT: vpbroadcastq 56(%r8), %ymm17
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm14, %zmm16, %zmm6
-; AVX512F-FAST-NEXT: vpbroadcastq 88(%r8), %ymm14
-; AVX512F-FAST-NEXT: vpbroadcastq 96(%r8), %ymm16
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm14, %zmm14
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm2, %zmm16, %zmm14
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm8, %zmm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm12, %zmm5
-; AVX512F-FAST-NEXT: vpbroadcastq 24(%r8), %ymm2
-; AVX512F-FAST-NEXT: vpbroadcastq 32(%r8), %ymm7
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm2, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm16, %zmm2
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm24, %zmm12, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm12, %zmm3
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm0, %zmm9, %zmm13
-; AVX512F-FAST-NEXT: vpternlogq $248, %zmm0, %zmm3, %zmm4
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, 384(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm13, 64(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 128(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 256(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, 448(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 576(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, (%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm27, 192(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm11, 320(%r9)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm29, 512(%r9)
-; AVX512F-FAST-NEXT: addq $456, %rsp # imm = 0x1C8
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpandn %ymm0, %ymm9, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm3
+; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm3, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm24
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm26
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8],ymm1[9],ymm0[10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm23[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3],ymm2[4],ymm1[5,6],ymm2[7],ymm1[8,9],ymm2[10],ymm1[11],ymm2[12],ymm1[13,14],ymm2[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [2,3,2,3,10,11,10,10]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm20, %zmm1
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm11 = [26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31,26,27,28,29,30,31,30,31]
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm18[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3,4],ymm0[5,6,7,8],ymm2[9],ymm0[10],ymm2[11,12],ymm0[13,14,15]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm6, %ymm2
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5],ymm3[6],ymm2[7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13],ymm3[14],ymm2[15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [2,3,2,3,10,10,11,10]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm18, %zmm2
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm31, %zmm2
+; AVX512F-FAST-NEXT: vpbroadcastq 112(%r8), %ymm0
+; AVX512F-FAST-NEXT: vpbroadcastq 120(%r8), %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm12
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm2, %zmm17, %zmm12
+; AVX512F-FAST-NEXT: vmovdqa 64(%rsi), %ymm6
+; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 64(%rdi), %ymm23
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm23[1,1,2,2]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5],ymm0[6],ymm1[7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm6, %ymm2
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3],ymm1[4],ymm2[5,6],ymm1[7],ymm2[8,9],ymm1[10],ymm2[11],ymm1[12],ymm2[13,14],ymm1[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm22
+; AVX512F-FAST-NEXT: vmovdqa 64(%rcx), %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm2
+; AVX512F-FAST-NEXT: vmovdqa 64(%rdx), %ymm3
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm3[3,0,3,0,7,4,7,4]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm8[0],ymm2[1],ymm8[2],ymm2[3],ymm8[4,5],ymm2[6],ymm8[7,8],ymm2[9],ymm8[10],ymm2[11],ymm8[12,13],ymm2[14],ymm8[15]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm8
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0],ymm15[1],ymm8[2,3],ymm15[4],ymm8[5],ymm15[6],ymm8[7,8],ymm15[9],ymm8[10,11],ymm15[12],ymm8[13],ymm15[14],ymm8[15]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm8
+; AVX512F-FAST-NEXT: vpbroadcastq 80(%r8), %ymm2
+; AVX512F-FAST-NEXT: vpandn %ymm2, %ymm9, %ymm9
+; AVX512F-FAST-NEXT: vmovdqa 64(%r8), %ymm5
+; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm15
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm15, %zmm15
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29,30,31,28,29,26,27,30,31,30,31,28,29,30,31,28,29]
+; AVX512F-FAST-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm9
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm30[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm14[1],ymm9[2],ymm14[3],ymm9[4,5],ymm14[6],ymm9[7,8],ymm14[9],ymm9[10],ymm14[11],ymm9[12,13],ymm14[14],ymm9[15]
+; AVX512F-FAST-NEXT: vprolq $16, %ymm13, %ymm13
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm30[0,1,2,1,4,5,6,5]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3],ymm14[4],ymm13[5,6],ymm14[7],ymm13[8,9],ymm14[10],ymm13[11],ymm14[12],ymm13[13,14],ymm14[15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm9, %zmm20, %zmm13
+; AVX512F-FAST-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm9
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm27[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm14[1],ymm9[2],ymm14[3,4],ymm9[5,6,7,8],ymm14[9],ymm9[10],ymm14[11,12],ymm9[13,14,15]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm0, %ymm7
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm27[1,1,1,2,5,5,5,6]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm14 = ymm7[0],ymm14[1],ymm7[2,3],ymm14[4],ymm7[5],ymm14[6],ymm7[7,8],ymm14[9],ymm7[10,11],ymm14[12],ymm7[13],ymm14[14],ymm7[15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm9, %zmm18, %zmm14
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm13, %zmm31, %zmm14
+; AVX512F-FAST-NEXT: vpbroadcastq 48(%r8), %ymm7
+; AVX512F-FAST-NEXT: vpbroadcastq 56(%r8), %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm14, %zmm17, %zmm7
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm9 = xmm9[0],mem[0],xmm9[1],mem[1],xmm9[2],mem[2],xmm9[3],mem[3]
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3,4],ymm1[5,6,7,8],ymm0[9],ymm1[10],ymm0[11,12],ymm1[13,14,15]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [4,5,2,3,2,3,0,1,10,11,8,9,4,5,6,7]
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm9, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [2,2,3,2,8,9,8,9]
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm9, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa 96(%rdi), %xmm1
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm6
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm23[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm13[1],ymm6[2],ymm13[3],ymm6[4,5],ymm13[6],ymm6[7,8],ymm13[9],ymm6[10],ymm13[11],ymm6[12,13],ymm13[14],ymm6[15]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,8,9,10,11,4,5,4,5,6,7,12,13]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [2,3,2,2,8,9,8,9]
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm13, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm1, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm4, %ymm4
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm5[0,1,1,1]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm10[3,2,3,3,7,6,7,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3,4],ymm4[5,6,7,8],ymm5[9],ymm4[10],ymm5[11,12],ymm4[13,14,15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm9, %zmm4
+; AVX512F-FAST-NEXT: vpbroadcastq 88(%r8), %ymm0
+; AVX512F-FAST-NEXT: vpbroadcastq 96(%r8), %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm6, %zmm5, %zmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm6
+; AVX512F-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
+; AVX512F-FAST-NEXT: # xmm6 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3]
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm6, %xmm6
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm17, %ymm9
+; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm3, %ymm9
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm24[0,1,1,1]
+; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm28[2,3,2,3,6,7,6,7]
+; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm11[1],ymm9[2],ymm11[3],ymm9[4,5],ymm11[6],ymm9[7,8],ymm11[9],ymm9[10],ymm11[11],ymm9[12,13],ymm11[14],ymm9[15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm6, %zmm13, %zmm9
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm1, %zmm9
+; AVX512F-FAST-NEXT: vpbroadcastq 24(%r8), %ymm4
+; AVX512F-FAST-NEXT: vpbroadcastq 32(%r8), %ymm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm9, %zmm5, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0,0,65535,65535,65535,0]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm3 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm16 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm5, %zmm3, %zmm9
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm5, %zmm16, %zmm29
+; AVX512F-FAST-NEXT: vpbroadcastq 64(%r8), %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm19, %zmm5, %zmm2
+; AVX512F-FAST-NEXT: vpbroadcastq (%r8), %ymm6
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm21, %zmm5, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm25 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm22, %zmm1, %zmm8
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm1, %zmm25, %zmm26
+; AVX512F-FAST-NEXT: vpternlogq $248, %zmm1, %zmm8, %zmm15
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, 128(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 256(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm0, 448(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm15, 384(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm12, 576(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm26, 64(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, (%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 320(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm29, 192(%r9)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, 512(%r9)
+; AVX512F-FAST-NEXT: addq $360, %rsp # imm = 0x168
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll
index a89f4a4b9d4571f..4522be032322c9a 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
@@ -3281,972 +3281,896 @@ define void @store_i16_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride6_vf32:
; AVX512F-ONLY-SLOW: # %bb.0:
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm2, %xmm24
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm1, %xmm23
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm31
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm9
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm3, %xmm25
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm2, %xmm28
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm12
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm5[0],ymm12[0],ymm5[1],ymm12[1],ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[8],ymm12[8],ymm5[9],ymm12[9],ymm5[10],ymm12[10],ymm5[11],ymm12[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm10
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm10[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm2
; AVX512F-ONLY-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm15
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm1, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6],ymm3[7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm3, %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm3, %xmm26
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm1, %xmm27
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm0[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm0[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm6, %xmm30
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm4, %xmm29
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm0[1,1,1,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm0[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm11
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm16 = xmm0[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm16[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm10
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm2, %zmm10 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm10, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1],ymm6[2],ymm9[3,4],ymm6[5],ymm9[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm9
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm6, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0],ymm13[1],ymm10[2,3],ymm13[4],ymm10[5,6],ymm13[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm10[0,1,2,3],zmm9[4,5,6,7]
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} ymm10 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm10 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm13[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm2, %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm2[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} ymm5 = 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
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[1,2,3,3,5,6,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm5[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm6, %zmm7 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm7, %ymm11
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm6, %ymm4
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm3, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0],ymm4[1],ymm9[2,3],ymm4[4],ymm9[5,6],ymm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm10[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm2[0,1,2,3],zmm9[4,5,6,7]
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm11[0],ymm4[1],ymm11[2,3],ymm4[4],ymm11[5,6],ymm4[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm7 = ymm11[0],ymm7[1,2],ymm11[3],ymm7[4,5],ymm11[6],ymm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm7[0,1,2,3],zmm4[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm7[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm7[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm5[4],ymm12[4],ymm5[5],ymm12[5],ymm5[6],ymm12[6],ymm5[7],ymm12[7],ymm5[12],ymm12[12],ymm5[13],ymm12[13],ymm5[14],ymm12[14],ymm5[15],ymm12[15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm31, %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm31[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm14[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm3[0],ymm9[1],ymm3[1],ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[8],ymm3[8],ymm9[9],ymm3[9],ymm9[10],ymm3[10],ymm9[11],ymm3[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm4[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm5[1,2,3,3,5,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm9[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm14[4],ymm10[4],ymm14[5],ymm10[5],ymm14[6],ymm10[6],ymm14[7],ymm10[7],ymm14[12],ymm10[12],ymm14[13],ymm10[13],ymm14[14],ymm10[14],ymm14[15],ymm10[15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm9[3,3,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm2, %zmm4 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm15, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm4, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm2[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm1[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm1[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm4 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm29, %xmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm30, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm29[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm30[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1,2],ymm4[3],ymm1[4,5],ymm4[6],ymm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm24 = <17,18,17,18,u,u,19,19,5,4,2,2,5,4,6,6>
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm7, %zmm24, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm11
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[1,1,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm3
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm5[2],ymm0[3,4],ymm5[5],ymm0[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm6 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm6, %xmm0, %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm7[1],ymm2[2,3],ymm7[4],ymm2[5,6],ymm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm5[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm7, %zmm24, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm12
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm4[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm4
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm4[4],xmm12[4],xmm4[5],xmm12[5],xmm4[6],xmm12[6],xmm4[7],xmm12[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm7, %zmm9 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm9, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm7[0,1],ymm1[2],ymm7[3,4],ymm1[5],ymm7[6,7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm4, %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm25, %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm28, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,0,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm6, %xmm7, %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1],ymm9[2,3],ymm6[4],ymm9[5,6],ymm6[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm6[0,1,2,3],zmm1[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm24, %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm23, %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm24[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm23[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm3, %zmm5, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm12[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm4[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm4, %zmm5 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm26, %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm8[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0],ymm4[1,2],ymm6[3],ymm4[4,5],ymm6[6],ymm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm26[0],zero,xmm26[1],zero,xmm26[2],zero,xmm26[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm3[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm25, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm9[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm6, %zmm4 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm4, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm7[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0],ymm6[1,2],ymm9[3],ymm6[4,5],ymm9[6],ymm6[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm7, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm6[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm10[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1],ymm6[2],ymm5[3,4],ymm6[5],ymm5[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm5[0,1,2,3],zmm4[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm27, %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm10[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm10 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm12 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm10, %zmm5, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm11[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm5[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm9, %zmm5 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm10 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1,2],ymm10[3],ymm9[4,5],ymm10[6],ymm9[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm0, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1],ymm0[2],ymm5[3,4],ymm0[5],ymm5[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm8
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm18, %zmm9, %zmm8
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm22, %zmm10
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm9, %zmm10
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm3, %zmm3
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm17, %zmm9, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm14, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm13, %zmm9, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm2, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm2, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 128(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 320(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm9
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm17, %zmm10, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm11
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm20, %zmm10, %zmm11
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm16, %zmm10
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm23, %zmm12, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm8, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm12, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm1
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm6, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm6, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, 192(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 256(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, 128(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rax)
; AVX512F-ONLY-SLOW-NEXT: vzeroupper
; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf32:
; AVX512F-ONLY-FAST: # %bb.0:
; AVX512F-ONLY-FAST-NEXT: pushq %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm13[0],ymm1[0],ymm13[1],ymm1[1],ymm13[2],ymm1[2],ymm13[3],ymm1[3],ymm13[8],ymm1[8],ymm13[9],ymm1[9],ymm13[10],ymm1[10],ymm13[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm0, %zmm20, %zmm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = 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]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm3
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm25
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [2,1,2,3,11,11,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm6, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm13
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [5,6,5,6,5,6,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm10, %ymm3
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
; AVX512F-ONLY-FAST-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm23 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [8,9,20,11,12,21,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm14[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm23, %zmm21
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm22, %zmm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm15[0],ymm2[1],ymm15[1],ymm2[2],ymm15[2],ymm2[3],ymm15[3],ymm2[8],ymm15[8],ymm2[9],ymm15[9],ymm2[10],ymm15[10],ymm2[11],ymm15[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm6, %xmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm4, %xmm18
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm10 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm10, %xmm3, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [0,9,2,3,8,5,6,11]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm16, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm1, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm11, %zmm20, %zmm25
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm17, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm6, %zmm25 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm6 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm6, %zmm25, %zmm22
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm10, %xmm6, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm10, %ymm16, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm10, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm20 = ymm4[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [5,6,5,6,5,6,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm2[4],ymm15[4],ymm2[5],ymm15[5],ymm2[6],ymm15[6],ymm2[7],ymm15[7],ymm2[12],ymm15[12],ymm2[13],ymm15[13],ymm2[14],ymm15[14],ymm2[15],ymm15[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm24, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm15 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm15[0],ymm2[1],ymm15[1],ymm2[2],ymm15[2],ymm2[3],ymm15[3],ymm2[8],ymm15[8],ymm2[9],ymm15[9],ymm2[10],ymm15[10],ymm2[11],ymm15[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm2[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm12[0],ymm4[0],ymm12[1],ymm4[1],ymm12[2],ymm4[2],ymm12[3],ymm4[3],ymm12[8],ymm4[8],ymm12[9],ymm4[9],ymm12[10],ymm4[10],ymm12[11],ymm4[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm16 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [8,21,10,11,20,13,14,23]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm11, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm15, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm26, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm28, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm13[4],ymm0[4],ymm13[5],ymm0[5],ymm13[6],ymm0[6],ymm13[7],ymm0[7],ymm13[12],ymm0[12],ymm13[13],ymm0[13],ymm13[14],ymm0[14],ymm13[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm24, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm12 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm12[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm12 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm13[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[8],ymm12[8],ymm13[9],ymm12[9],ymm13[10],ymm12[10],ymm13[11],ymm12[11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm9, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm2[0],ymm13[0],ymm2[1],ymm13[1],ymm2[2],ymm13[2],ymm2[3],ymm13[3],ymm2[8],ymm13[8],ymm2[9],ymm13[9],ymm2[10],ymm13[10],ymm2[11],ymm13[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = 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]
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512F-ONLY-FAST-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm11, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm9[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm12, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm5 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm14, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm0, %zmm5, %zmm15
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm11, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm1[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm14[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm26, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm1[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm18, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm19, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm8, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm7, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
-; AVX512F-ONLY-FAST-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm4, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm7[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [8,21,10,11,20,13,14,23]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm2
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm11, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [12,1,2,13,4,5,14,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm17, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm1[0,1,2,3],zmm4[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm3
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm3, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
+; AVX512F-ONLY-FAST-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm3, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = <2,2,u,3,10,u,10,11>
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm18, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm9, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm28
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm6, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[12],ymm1[12],ymm6[13],ymm1[13],ymm6[14],ymm1[14],ymm6[15],ymm1[15]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm10, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm10 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm10[0],ymm4[1],ymm10[1],ymm4[2],ymm10[2],ymm4[3],ymm10[3],ymm4[8],ymm10[8],ymm4[9],ymm10[9],ymm4[10],ymm10[10],ymm4[11],ymm10[11]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm10, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm2, %zmm0, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm17, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm0[0,1,2,3],zmm11[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm11, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm11, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm18, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[8],ymm1[8],ymm6[9],ymm1[9],ymm6[10],ymm1[10],ymm6[11],ymm1[11]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm12
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,0,2,1,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm21, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm23 = [1,0,2,2,1,0,2,2]
+; AVX512F-ONLY-FAST-NEXT: # ymm23 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm23, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm1
; AVX512F-ONLY-FAST-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [16,9,10,17,12,13,18,15]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm6[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm3, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm10[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm4, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm10, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm11, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm9, %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm14[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm6, %zmm8 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm27, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm6, %zmm8, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm6 = xmm27[0],zero,xmm27[1],zero,xmm27[2],zero,xmm27[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm9[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm10, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm9, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm12 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [16,9,10,17,12,13,18,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm8 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm8, %zmm5, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [0,1,8,3,4,9,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm8, %ymm22, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm12[0,1,2,3],zmm1[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm8, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm21, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = <u,0,0,u,8,8,u,9>
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm7, %xmm12, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm12[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm21, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm23, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm14 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm14, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm4, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm4, %zmm2 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm14, %zmm2, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm14 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm14, %ymm22, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm2[0,1,2,3],zmm5[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm7, %xmm5, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm21, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm28, %ymm2
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm15 = [1,1,1,1,10,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm15, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm18, %zmm7, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm18 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [8,9,20,11,12,21,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm8 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm8, %zmm0, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [0,9,2,3,8,5,6,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm21, %ymm18
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm5, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = <0,u,0,1,u,10,10,u>
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm5, %zmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm24, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm25, %ymm9
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm13[4],xmm3[5],xmm13[5],xmm3[6],xmm13[6],xmm3[7],xmm13[7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm15, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm20, %zmm7, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm4 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm3, %zmm4, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm6, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm21, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm12, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm5, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm0[0,1,2,3]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm23[0,1,2,3],zmm21[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm9, %zmm18
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm14, %zmm19, %zmm18
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm25[0,1,2,3],zmm22[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm14, %zmm19, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm16[0,1,2,3],zmm17[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm28, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm14, %zmm11, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm15[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm26, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm5, %zmm11, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [0,1,8,3,4,9,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm5, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm12[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm2
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm7, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm6, %ymm5, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm8[0,1,2,3],zmm3[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm10, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm7, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 64(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm18[0,1,2,3],zmm10[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm22, %zmm0, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, (%rax)
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm19, %zmm0, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 192(%rax)
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm17, %zmm0, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm16, %zmm0, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 320(%rax)
; AVX512F-ONLY-FAST-NEXT: popq %rax
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride6_vf32:
; AVX512DQ-SLOW: # %bb.0:
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm2, %xmm23
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm1, %xmm24
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm14
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm2[0],ymm14[1],ymm2[1],ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[8],ymm2[8],ymm14[9],ymm2[9],ymm14[10],ymm2[10],ymm14[11],ymm2[11]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm31
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm3
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm28
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm2, %xmm27
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm12
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm5[0],ymm12[0],ymm5[1],ymm12[1],ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[8],ymm12[8],ymm5[9],ymm12[9],ymm5[10],ymm12[10],ymm5[11],ymm12[11]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
-; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm15
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm15[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm3
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm3, %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm25
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm1
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm1, %xmm26
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm0[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm7
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm0[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm27
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm2, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm4
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm5
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm5, %xmm28
; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm29
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm4
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %ymm3
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = 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-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm0[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm8
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm11[4],xmm8[5],xmm11[5],xmm8[6],xmm11[6],xmm8[7],xmm11[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm0[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm10 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1,2],ymm3[3],ymm0[4,5],ymm3[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm31
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm0, %ymm26
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm1, %xmm30
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm0[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3]
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm6, %xmm25
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm3 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm2, %zmm0
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm12[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm6
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm10, %zmm16, %zmm6 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm9
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1],ymm10[2],ymm9[3,4],ymm10[5],ymm9[6,7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm10
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm10, %xmm13
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm13[1],ymm6[2,3],ymm13[4],ymm6[5,6],ymm13[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm6[0,1,2,3],zmm9[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm1[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm0[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[8],ymm6[8],ymm9[9],ymm6[9],ymm9[10],ymm6[10],ymm9[11],ymm6[11]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm0 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm9 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm9 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[2],ymm0[2],ymm9[3],ymm0[3],ymm9[8],ymm0[8],ymm9[9],ymm0[9],ymm9[10],ymm0[10],ymm9[11],ymm0[11]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm13[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm5, %xmm17
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm4[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm0, %ymm22
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm4[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm16
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm0[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm8
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm7
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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-SLOW-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm6, %zmm4
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm9, %zmm4 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm6
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm10, %ymm6, %ymm1
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm4, %ymm6
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm1[1],ymm6[2,3],ymm1[4],ymm6[5,6],ymm1[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm9[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm17 = zmm2[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm5[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[8],ymm2[8],ymm4[9],ymm2[9],ymm4[10],ymm2[10],ymm4[11],ymm2[11]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm5[4],ymm12[4],ymm5[5],ymm12[5],ymm5[6],ymm12[6],ymm5[7],ymm12[7],ymm5[12],ymm12[12],ymm5[13],ymm12[13],ymm5[14],ymm12[14],ymm5[15],ymm12[15]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm31, %ymm9
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm2 = 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-SLOW-NEXT: vpsrldq {{.*#+}} ymm6 = 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-SLOW-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-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm4[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm5[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm6[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm14[4],ymm9[4],ymm14[5],ymm9[5],ymm14[6],ymm9[6],ymm14[7],ymm9[7],ymm14[12],ymm9[12],ymm14[13],ymm9[13],ymm14[14],ymm9[14],ymm14[15],ymm9[15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[1,2,3,3,5,6,7,7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm6[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm4, %zmm5, %zmm1 {%k1}
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm15, %ymm3
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm4
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm15[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm4
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm0[3,3,3,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa (%rdx), %ymm1
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm9, %zmm2, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm10, %ymm2, %ymm9
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm10
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6],ymm9[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm10[0],ymm0[1,2],ymm10[3],ymm0[4,5],ymm10[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm0[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm9
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm9[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm0[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm25, %xmm0
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm0[4],xmm11[5],xmm0[5],xmm11[6],xmm0[6],xmm11[7],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm10[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = <17,18,17,18,u,u,19,19,5,4,2,2,5,4,6,6>
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm0, %zmm10, %zmm3
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm12[4],xmm15[5],xmm12[5],xmm15[6],xmm12[6],xmm15[7],xmm12[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm1[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm5[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1,2],ymm4[3],ymm1[4,5],ymm4[6],ymm1[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm13 = zmm1[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm1[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm8[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm4, %zmm0, %zmm3 {%k1}
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm17, %xmm4
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3],ymm4[4],ymm3[5,6],ymm4[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm16[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm29, %xmm6
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm30, %xmm7
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,0,2,2]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm4[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm4 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm6 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm4, %ymm4
-; AVX512DQ-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm4, %zmm3, %zmm1 {%k1}
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm3
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm10[2,1,3,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1,2],ymm4[3],ymm3[4,5],ymm4[6],ymm3[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm4, %ymm4
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm3[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm28, %xmm6
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm27, %xmm7
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm28[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm27[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm23, %xmm8
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm24, %xmm9
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,0,2,2]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm7[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm7 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm7, %zmm6, %zmm4 {%k1}
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm4, %ymm6
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm25, %xmm8
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm8[2,1,3,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1,2],ymm7[3],ymm6[4,5],ymm7[6],ymm6[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm7 = xmm25[0],zero,xmm25[1],zero,xmm25[2],zero,xmm25[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm26, %xmm7
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm27, %xmm5
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm4 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm4 = xmm5[4],mem[4],xmm5[5],mem[5],xmm5[6],mem[6],xmm5[7],mem[7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermt2d %zmm4, %zmm10, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm28, %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm29, %xmm10
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm10[4],xmm4[5],xmm10[5],xmm4[6],xmm10[6],xmm4[7],xmm10[7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm7, %zmm4, %zmm5 {%k1}
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm4
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm6[2],ymm4[3,4],ymm6[5],ymm4[6,7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm31, %xmm6
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm6, %xmm0
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2,3],ymm0[4],ymm5[5,6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm30[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm19, %zmm8
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm18, %zmm9, %zmm8
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm22, %zmm10
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm9, %zmm10
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm2, %zmm2
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm26, %zmm6
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm22, %zmm8
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm7, %zmm8
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm7
; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm17, %zmm9, %zmm2
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm14, %zmm5
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm13, %zmm9, %zmm5
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm1
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm3, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 320(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm20, %zmm9, %zmm7
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm10
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm23, %zmm9, %zmm10
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm3, %zmm1
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm2
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, 192(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, 320(%rax)
; AVX512DQ-SLOW-NEXT: vzeroupper
; AVX512DQ-SLOW-NEXT: retq
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf32:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $40, %rsp
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm22
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm4
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm5, %ymm1
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm3, %ymm29
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm3 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm17
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm1
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm1, %ymm25
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm20, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm3, %ymm27
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512DQ-FAST-NEXT: vpermd %zmm2, %zmm21, %zmm19
; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm2, %zmm22 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm1, %zmm19 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm24
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, %zmm23
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm17, %zmm23
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm4
+; AVX512DQ-FAST-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-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm6, %xmm30
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm5, %xmm29
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm0, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm28
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm6, %ymm19
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = <0,u,0,1,u,10,10,u>
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm12
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm18, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512DQ-FAST-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-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm20, %zmm0
+; AVX512DQ-FAST-NEXT: vpermd %zmm3, %zmm21, %zmm20
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}
; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm0, %zmm20, %zmm17
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm11
+; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm11, %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm6, %ymm20
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm13
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm31
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm18, %zmm13
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm15
+; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm21 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FAST-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]
+; AVX512DQ-FAST-NEXT: vpermd %ymm1, %ymm21, %ymm1
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm9 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} 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-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm1, %zmm4, %zmm0 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm9
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, %zmm20
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm2, %zmm21, %zmm20
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm3, %ymm30
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm2, %ymm9, %ymm22
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm6
-; AVX512DQ-FAST-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-FAST-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm3
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} 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-FAST-NEXT: vpshufb %ymm3, %ymm6, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm6, %ymm31
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm10
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm10, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm8
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[8],ymm4[8],ymm0[9],ymm4[9],ymm0[10],ymm4[10],ymm0[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = 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]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm23
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm13
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm4[4],ymm13[4],ymm4[5],ymm13[5],ymm4[6],ymm13[6],ymm4[7],ymm13[7],ymm4[12],ymm13[12],ymm4[13],ymm13[13],ymm4[14],ymm13[14],ymm4[15],ymm13[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm5, %ymm0
-; AVX512DQ-FAST-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
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm12 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm12[0],ymm5[0],ymm12[1],ymm5[1],ymm12[2],ymm5[2],ymm12[3],ymm5[3],ymm12[8],ymm5[8],ymm12[9],ymm5[9],ymm12[10],ymm5[10],ymm12[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm5, %zmm23 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm12
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm0
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm0, %zmm23, %zmm21
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm12[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm9, %ymm23
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm1
-; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm15
; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm0[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm4[0],ymm13[0],ymm4[1],ymm13[1],ymm4[2],ymm13[2],ymm4[3],ymm13[3],ymm4[8],ymm13[8],ymm4[9],ymm13[9],ymm4[10],ymm13[10],ymm4[11],ymm13[11]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm27
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm2, %xmm4
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm13
-; AVX512DQ-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm25 = [1,0,2,2,1,0,2,2]
-; AVX512DQ-FAST-NEXT: # ymm25 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm25, %ymm7
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm0 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm11 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm7, %zmm4 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm5
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm26, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm11
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm11, %xmm3
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm25, %ymm28
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm3
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm14 = xmm3[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm14, %ymm25
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm15, %zmm4, %zmm9
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm22, %ymm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm18 = zmm0[0,1,2,3],zmm9[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm14, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm25, %ymm9
+; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm9, %ymm3
+; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vpermt2q %zmm3, %zmm2, %zmm0
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm6
+; AVX512DQ-FAST-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-FAST-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm9, %ymm5, %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = <2,2,u,3,10,u,10,11>
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm5, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm27, %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm26, %ymm1
+; AVX512DQ-FAST-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-FAST-NEXT: vpermd %ymm6, %ymm21, %ymm6
+; AVX512DQ-FAST-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-FAST-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-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm14, %zmm0 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm24, %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm6, %ymm1
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm1, %zmm0, %zmm4
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm22, %ymm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm0[0,1,2,3],zmm4[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm7, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufb %ymm9, %ymm7, %ymm0
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm5, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm10, %xmm3
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm8, %xmm4
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,2,1,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm5, %zmm3
+; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm29, %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm30, %xmm9
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm6, %ymm4, %ymm6
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm7[0],xmm11[1],xmm7[1],xmm11[2],xmm7[2],xmm11[3],xmm7[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm28, %zmm0 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm9
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm6, %zmm0, %zmm26
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = 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]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
-; AVX512DQ-FAST-NEXT: vpermd %zmm27, %zmm10, %zmm8
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm6, %zmm8 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm12[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm12
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm7, %zmm6, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm29, %ymm7
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm17, %ymm11
-; AVX512DQ-FAST-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]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm13[4],xmm1[5],xmm13[5],xmm1[6],xmm13[6],xmm1[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm7
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm9, %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm1, %xmm17
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm13, %ymm27, %ymm8
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm3, %xmm13
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm13[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero
-; AVX512DQ-FAST-NEXT: vpermd %zmm7, %zmm10, %zmm7
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm10
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm15 = xmm10[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm15, %ymm15
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm10, %xmm11
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm14 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm10, %xmm10
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm13
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm14, %zmm2, %zmm7 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm30, %ymm2
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm2, %zmm7, %zmm6
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm14
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm5, %xmm5
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm27, %ymm7
-; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm1
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm3
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm22[0,1,2,3],zmm20[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm14 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm16, %zmm14
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm23[0,1,2,3],zmm21[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm17
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm16, %zmm17
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm6, %zmm3 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm8 = xmm11[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm9
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm8, %zmm6, %zmm9
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm11, %xmm10
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm7, %xmm14, %xmm7
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm7[4],xmm10[4],xmm7[5],xmm10[5],xmm7[6],xmm10[6],xmm7[7],xmm10[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm11[0],xmm14[1],xmm11[1],xmm14[2],xmm11[2],xmm14[3],xmm11[3]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm5, %zmm10
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm2, %ymm5, %ymm4
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm9, %ymm5, %ymm0
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm4[0,1,2,3],zmm24[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm25, %zmm4
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm8, %ymm5, %ymm3
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm9[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm4, %ymm4
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm7 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm7, %ymm7
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm4, %zmm10 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm31, %xmm7
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm7, %xmm4
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,0,0,u,8,8,u,9>
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm9, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm28, %xmm11
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm4, %zmm10, %zmm6
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm28[0],zero,xmm28[1],zero,xmm28[2],zero,xmm28[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm4, %ymm5, %ymm10
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm10[0,1,2,3],zmm6[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm6, %xmm5
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm5, %zmm9, %zmm6
; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm5, %zmm4
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm26[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm2
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm12[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm10, %zmm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm5
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm7[0,1,2,3],zmm6[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm8, %zmm1
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm5, %zmm6
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 64(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512DQ-FAST-NEXT: addq $40, %rsp
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, (%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm5, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $184, %zmm1, %zmm3, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, 128(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm18, %zmm3, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 320(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm20[0,1,2,3],zmm17[0,1,2,3]
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm13
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, 256(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm19[0,1,2,3],zmm23[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, 64(%rax)
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
@@ -7496,1942 +7420,1878 @@ define void @store_i16_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512F-ONLY-SLOW-LABEL: store_i16_stride6_vf64:
; AVX512F-ONLY-SLOW: # %bb.0:
-; AVX512F-ONLY-SLOW-NEXT: subq $392, %rsp # imm = 0x188
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512F-ONLY-SLOW-NEXT: subq $600, %rsp # imm = 0x258
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %ymm9
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %ymm6
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm4, %ymm19
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %ymm10
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %ymm5
+; AVX512F-ONLY-SLOW-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>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm5, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm3, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm0[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %xmm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[1,1,1,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %ymm3
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm0[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm6[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm11[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm4, %zmm6 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm4[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm11
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm11[0,1],ymm5[2],ymm11[3,4],ymm5[5],ymm11[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm5, %xmm12
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm13, %xmm16
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm12[1],ymm6[2,3],ymm12[4],ymm6[5,6],ymm12[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm6[0,1,2,3],zmm11[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %xmm11
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm11[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm11, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm6[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm11[2,2,2,2]
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm9 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm2[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm3[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[8],ymm10[8],ymm12[9],ymm10[9],ymm12[10],ymm10[10],ymm12[11],ymm10[11]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm9[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm3, %zmm2 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm4, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm9, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm9[0],ymm3[1],ymm9[2,3],ymm3[4],ymm9[5,6],ymm3[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0],ymm2[1,2],ymm4[3],ymm2[4,5],ymm4[6],ymm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm3[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} ymm2 = ymm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm10[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %ymm9
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[1,2,3,3,5,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm2[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm12 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[3,3,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm3, %ymm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm14, %ymm31
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm12
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2,3],ymm11[4],ymm12[5,6],ymm11[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm12[0],ymm6[1,2],ymm12[3],ymm6[4,5],ymm12[6],ymm6[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm11[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm11[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm6[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm11[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,0,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm11[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k2
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k2}
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1,2],ymm0[3],ymm7[4,5],ymm0[6],ymm7[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm5[2],ymm1[3,4],ymm5[5],ymm1[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm13[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm13, %xmm17
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm0, %ymm24
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,0,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm7 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm8 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm7, %ymm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %xmm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm7[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm8[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm5, %zmm11 {%k2}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm11, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0],ymm14[1,2],ymm12[3],ymm14[4,5],ymm12[6],ymm14[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm12
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm14 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm14, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm14[2],ymm11[3,4],ymm14[5],ymm11[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm11[0,1,2,3],zmm12[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[1,1,1,1]
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[8],ymm10[8],ymm9[9],ymm10[9],ymm9[10],ymm10[10],ymm9[11],ymm10[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm7, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %ymm2
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %ymm3
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm3[4],ymm2[4],ymm3[5],ymm2[5],ymm3[6],ymm2[6],ymm3[7],ymm2[7],ymm3[12],ymm2[12],ymm3[13],ymm2[13],ymm3[14],ymm2[14],ymm3[15],ymm2[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm21
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm16, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm5, %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %xmm7
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm7[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm0[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm0[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm0[1,1,1,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm0[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm0[1,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[1,0,2,2,5,4,6,6]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm14[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm15[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm14
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm14, %ymm15
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm15[0,1],ymm3[2],ymm15[3,4],ymm3[5],ymm15[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm15
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm3, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3],ymm13[4],ymm14[5,6],ymm13[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm13[0,1,2,3],zmm15[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} ymm13 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} ymm14 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm8[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm9[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm14[0],ymm1[0],ymm14[1],ymm1[1],ymm14[2],ymm1[2],ymm14[3],ymm1[3],ymm14[8],ymm1[8],ymm14[9],ymm1[9],ymm14[10],ymm1[10],ymm14[11],ymm1[11]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[12],ymm8[12],ymm9[13],ymm8[13],ymm9[14],ymm8[14],ymm9[15],ymm8[15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[1,2,3,3,5,6,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm13, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm4, %ymm16
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm31, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm10, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2,3],ymm0[4],ymm8[5,6],ymm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %ymm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm11, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm18, %ymm9
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm19, %ymm8
-; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512F-ONLY-SLOW-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
; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm1 = 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]
+; AVX512F-ONLY-SLOW-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]
; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,2,3,3,5,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm13[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm8[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm10[2,1,2,3,6,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm9 = 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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm14, %ymm23
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm4, %ymm24
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm1
; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm9, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm3, %ymm0
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0],ymm0[1],ymm8[2,3],ymm0[4],ymm8[5,6],ymm0[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0],ymm1[1,2],ymm8[3],ymm1[4,5],ymm8[6],ymm1[7]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm29, %zmm9
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm9 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm10
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm23, %zmm8, %zmm10
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm17, %xmm12
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm12[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm1[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm4 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm4, %ymm25
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,0,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm2[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm7[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm0[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm14[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm6, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm1, %zmm6 {%k2}
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm17
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm5[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm2
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm1
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm20 = ymm3[1,2,3,3,5,6,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm25, %zmm20
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm0, %zmm15
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm20, %zmm15 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm0, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm15, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm7 = ymm14[0],ymm7[1],ymm14[2,3],ymm7[4],ymm14[5,6],ymm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0],ymm15[1,2],ymm14[3],ymm15[4,5],ymm14[6],ymm15[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm14[0,1,2,3],zmm7[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm29 = <17,18,17,18,u,u,19,19,5,4,2,2,5,4,6,6>
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm14[4],xmm7[4],xmm14[5],xmm7[5],xmm14[6],xmm7[6],xmm14[7],xmm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm14, %xmm26
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm2, %zmm29, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm14[4],xmm2[5],xmm14[5],xmm2[6],xmm14[6],xmm2[7],xmm14[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1],ymm0[2],ymm2[3,4],ymm0[5],ymm2[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm2, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm3, %xmm14
; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3],ymm2[4],ymm1[5,6],ymm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm0, %zmm29, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm2, %zmm6 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm5[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm2, %xmm8
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0],ymm8[1],ymm6[2,3],ymm8[4],ymm6[5,6],ymm8[7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm6[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm4 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm31 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm3, %zmm31, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
+; AVX512F-ONLY-SLOW-NEXT: kmovw %eax, %k2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm5, %zmm0 {%k2}
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3],ymm3[4,5],ymm1[6],ymm3[7]
; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1],ymm3[2],ymm6[3,4],ymm3[5],ymm6[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm6[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm7 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm14 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm14[0],xmm7[0],xmm14[1],xmm7[1],xmm14[2],xmm7[2],xmm14[3],xmm7[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,0,2,2]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm0[0,1,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm4[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm0 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm0, %zmm31, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm0[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm5, %zmm6 {%k2}
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm8[1,2],ymm5[3],ymm8[4,5],ymm5[6],ymm8[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1],ymm8[2],ymm6[3,4],ymm8[5],ymm6[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm6[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm21, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm19, %ymm5
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm2, %zmm29, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm18, %ymm4
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm1, %zmm3 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm14, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm0, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6],ymm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm0
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm1[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm0[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm1[2,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm11[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm0[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm13
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm1, %zmm29, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm23, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm24, %ymm3
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm8, %zmm6 {%k1}
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm17, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm14 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm4, %xmm8, %xmm9
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm6[0],ymm9[1],ymm6[2,3],ymm9[4],ymm6[5,6],ymm9[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm6 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm6 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm23 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm23 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm2, %zmm12
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm14
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm9[0,1,2,3],zmm14[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm13 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm0, %zmm31, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm15[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm1[0,1,2,1]
; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,7,6,5]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm13[4],xmm15[4],xmm13[5],xmm15[5],xmm13[6],xmm15[6],xmm13[7],xmm15[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}
-; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm7, %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm13[0],ymm6[1,2],ymm13[3],ymm6[4,5],ymm13[6],ymm6[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm4[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm23 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm23, %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %xmm13
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm13[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm14[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm14 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm14[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm9, %zmm0 {%k2}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm9[0],ymm1[1,2],ymm9[3],ymm1[4,5],ymm9[6],ymm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm9, %ymm18
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm13[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm9[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero,xmm8[2],zero,xmm8[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm8, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm8[2],ymm0[3,4],ymm8[5],ymm0[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %xmm8
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm26, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpsrldq {{.*#+}} xmm7 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm8[0,0,2,1,4,5,6,7]
; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm7, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm6[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm0[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm30, %zmm8, %zmm11
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm16, %zmm8, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm8 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm27, %zmm16
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm16 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm15, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm31, %zmm14, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm19, %zmm14, %zmm12
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm24, %zmm13
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm14, %zmm13
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm15
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm28, %zmm14, %zmm15
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm14, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm6, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm14, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm1, %zmm31, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm2[0,1,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm13 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm14[0,1,2,3],zmm1[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa32 %zmm0, %zmm1 {%k2}
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm3[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm14[0],ymm0[1,2],ymm14[3],ymm0[4,5],ymm14[6],ymm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm15[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,2,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm2 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm26 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm26, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm15[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm11[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,2,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm12 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm29, %zmm4 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm5 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm24, %zmm16
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm27, %zmm10, %zmm16
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm25, %zmm10, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm14, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm22, %zmm10, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm18, %zmm10
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm28, %zmm13, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm30, %zmm13, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm13, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm13, %zmm1
; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, 192(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, 128(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 320(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 192(%rax)
; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, 448(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm15, 384(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, 576(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, 512(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 704(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 640(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 64(%rax)
-; AVX512F-ONLY-SLOW-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 448(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 384(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, 576(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, 640(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, 128(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, 320(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm23, 512(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, 704(%rax)
+; AVX512F-ONLY-SLOW-NEXT: addq $600, %rsp # imm = 0x258
; AVX512F-ONLY-SLOW-NEXT: vzeroupper
; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride6_vf64:
; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $1272, %rsp # imm = 0x4F8
+; AVX512F-ONLY-FAST-NEXT: subq $1256, %rsp # imm = 0x4E8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm7, (%rsp) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0],ymm14[0],ymm6[1],ymm14[1],ymm6[2],ymm14[2],ymm6[3],ymm14[3],ymm6[8],ymm14[8],ymm6[9],ymm14[9],ymm6[10],ymm14[10],ymm6[11],ymm14[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm2[4],xmm13[5],xmm2[5],xmm13[6],xmm2[6],xmm13[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm2, %xmm26
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm9[0],ymm3[0],ymm9[1],ymm3[1],ymm9[2],ymm3[2],ymm9[3],ymm3[3],ymm9[8],ymm3[8],ymm9[9],ymm3[9],ymm9[10],ymm3[10],ymm9[11],ymm3[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm8, %xmm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm7, %xmm18
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm5, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[2],ymm0[2],ymm7[3],ymm0[3],ymm7[8],ymm0[8],ymm7[9],ymm0[9],ymm7[10],ymm0[10],ymm7[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm1
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm10, %xmm19
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm6, %xmm23
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm10
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm12, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm4, %zmm19, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm8
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [2,1,2,3,11,11,11,11]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm27, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [5,6,5,6,5,6,7,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm26, %ymm8
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm3, %zmm3
; AVX512F-ONLY-FAST-NEXT: movw $18724, %ax # imm = 0x4924
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm4, %zmm16 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [8,9,20,11,12,21,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm1
+; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm0 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [8,21,10,11,20,13,14,23]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm1
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm17, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [0,9,2,3,8,5,6,11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm27, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm4[0,1,0,1]
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm1, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm13, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm29 = [12,1,2,13,4,5,14,7]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm29, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm4[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm5[4],xmm11[5],xmm5[5],xmm11[6],xmm5[6],xmm11[7],xmm5[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = 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]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm20, %zmm19, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm20 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = <2,2,u,3,10,u,10,11>
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm0
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: # ymm14 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm28, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm3
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm0
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm26, %ymm3
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: 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]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm0 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm1, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm29, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[0,1,2,3]
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm15, %xmm1, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm27, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm28, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm4[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [5,6,5,6,5,6,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm4 = ymm6[4],ymm14[4],ymm6[5],ymm14[5],ymm6[6],ymm14[6],ymm6[7],ymm14[7],ymm6[12],ymm14[12],ymm6[13],ymm14[13],ymm6[14],ymm14[14],ymm6[15],ymm14[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm4, %ymm12, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm12, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm14[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} 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]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm4[2,2,2,2]
-; AVX512F-ONLY-FAST-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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm8, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm10, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm12 = ymm12[0],ymm14[0],ymm12[1],ymm14[1],ymm12[2],ymm14[2],ymm12[3],ymm14[3],ymm12[8],ymm14[8],ymm12[9],ymm14[9],ymm12[10],ymm14[10],ymm12[11],ymm14[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = 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]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm6, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm12, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm6, %zmm24 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm29 = [8,21,10,11,20,13,14,23]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm6, %zmm29, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm2, %ymm6, %ymm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm6, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm9[4],ymm3[4],ymm9[5],ymm3[5],ymm9[6],ymm3[6],ymm9[7],ymm3[7],ymm9[12],ymm3[12],ymm9[13],ymm3[13],ymm9[14],ymm3[14],ymm9[15],ymm3[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm25, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm2, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[8],ymm8[8],ymm9[9],ymm8[9],ymm9[10],ymm8[10],ymm9[11],ymm8[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm9 = ymm2[4],ymm0[4],ymm2[5],ymm0[5],ymm2[6],ymm0[6],ymm2[7],ymm0[7],ymm2[12],ymm0[12],ymm2[13],ymm0[13],ymm2[14],ymm0[14],ymm2[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm28 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm28, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm29, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm12, %ymm28
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm1, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [0,0,2,1,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
; AVX512F-ONLY-FAST-NEXT: vbroadcasti32x4 {{.*#+}} ymm31 = [1,0,2,2,1,0,2,2]
; AVX512F-ONLY-FAST-NEXT: # ymm31 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm26, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm31, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm8 = xmm13[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm5, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm11, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm5[0],xmm11[1],xmm5[1],xmm11[2],xmm5[2],xmm11[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm2 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512F-ONLY-FAST-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm30 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [16,9,10,17,12,13,18,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm30, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm26, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm18, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm23, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm31, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm8, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm11, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm13[4],xmm6[4],xmm13[5],xmm6[5],xmm13[6],xmm6[6],xmm13[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm13, %zmm23
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm23 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm23, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm26, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm3 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm3 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm22
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm17, %zmm22
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm5, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm27, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm0[1,1,1,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm8[0],ymm13[0],ymm8[1],ymm13[1],ymm8[2],ymm13[2],ymm8[3],ymm13[3],ymm8[8],ymm13[8],ymm8[9],ymm13[9],ymm8[10],ymm13[10],ymm8[11],ymm13[11]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm0[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm21 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm11, %zmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm11, %zmm21 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm11[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm1, %zmm21, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm0
+; AVX512F-ONLY-FAST-NEXT: kmovw %eax, %k2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm4 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = [16,9,10,17,12,13,18,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm0, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm27, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm7[4],ymm0[4],ymm7[5],ymm0[5],ymm7[6],ymm0[6],ymm7[7],ymm0[7],ymm7[12],ymm0[12],ymm7[13],ymm0[13],ymm7[14],ymm0[14],ymm7[15],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm7 = ymm7[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm7[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm7[0],ymm0[0],ymm7[1],ymm0[1],ymm7[2],ymm0[2],ymm7[3],ymm0[3],ymm7[8],ymm0[8],ymm7[9],ymm0[9],ymm7[10],ymm0[10],ymm7[11],ymm0[11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm8, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = 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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,20,21,20,21,22,23,24,25,24,25,24,25,24,25]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm14[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm8[4],ymm13[4],ymm8[5],ymm13[5],ymm8[6],ymm13[6],ymm8[7],ymm13[7],ymm8[12],ymm13[12],ymm8[13],ymm13[13],ymm8[14],ymm13[14],ymm8[15],ymm13[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm13[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm25, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm27
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm27 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm27, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm11, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm29, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm3, %zmm30, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm3, %ymm16, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm4[0,1,2,3],zmm2[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = <u,0,0,u,8,8,u,9>
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm1, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm20
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm18, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm23, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm19, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm2, %ymm31, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm4, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm30, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm24
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm16, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm2[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[8],ymm2[8],ymm3[9],ymm2[9],ymm3[10],ymm2[10],ymm3[11],ymm2[11]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm23
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm2
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm26, %ymm3
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm2 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm8, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm4, %zmm25, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm0[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm29, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm3[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm9, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm8, %ymm1
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm2, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm28, %zmm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm10[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm7[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm11[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = [12,1,2,13,4,5,14,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm7, %ymm18, %ymm27
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm7, %ymm25, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm3, %ymm26, %ymm3
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm3, %zmm1 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm4, %zmm1, %zmm25
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm4, %ymm29, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm1[0,1,2,3],zmm25[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm5, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm28, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm11, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm29
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm18, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm31, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm6, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm7 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm15 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm15, %zmm30, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm15 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm15, %ymm16, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm26 = zmm7[0,1,2,3],zmm1[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm15
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm15, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm18, %zmm12
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm13[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm13 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} ymm14 = ymm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm0[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm4[0],ymm14[0],ymm4[1],ymm14[1],ymm4[2],ymm14[2],ymm4[3],ymm14[3],ymm4[8],ymm14[8],ymm4[9],ymm14[9],ymm4[10],ymm14[10],ymm4[11],ymm14[11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} ymm1 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm14[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm11
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm11, %zmm4 {%k2}
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm10, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm1, %zmm4, %zmm29
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm18, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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]
-; AVX512F-ONLY-FAST-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm0, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm31, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm7, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm7[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm17, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm14[0],xmm2[1],xmm14[1],xmm2[2],xmm14[2],xmm2[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm1, %ymm31, %ymm18
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm14 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3]
; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm5, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm18, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm12 {%k2}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm14
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm14[2,1,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm1, %zmm12, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm16, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm12[0,1,2,3],zmm30[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm1, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm1[0,0,2,1,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm17, %zmm12
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [1,1,1,1,10,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm9, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm15 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm2 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm5[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm26, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm31, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm0 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpsrldq {{.*#+}} xmm1 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm16[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm16 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm20, %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm20[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm20 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm22[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm21[0,1,2,3],zmm17[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm9, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm24[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm18 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm18
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm28, %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = zmm28[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm15 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm15
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm27[0,1,2,3],zmm6[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm11[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm13[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm6, %zmm31, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm21 # 16-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # xmm21 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm9, %xmm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm6, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm11 = xmm12[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm11, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm12, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm22 # 16-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # xmm22 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm6, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm8[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm13, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = [8,9,20,11,12,21,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, %zmm17
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm18, %zmm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm14, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [0,9,2,3,8,5,6,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm14, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm1, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm30
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = <0,u,0,1,u,10,10,u>
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm1 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm1 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm9, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm5 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm1, %zmm5 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm1 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, %zmm28
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm1, %zmm18, %zmm28
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm1, %ymm14, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm20, %xmm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm9, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm1 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm1 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %zmm0, %zmm18, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm0, %ymm14, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm23, %ymm13
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm13
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm13[4],xmm11[5],xmm13[5],xmm11[6],xmm13[6],xmm11[7],xmm13[7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm9, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm0 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm11, %zmm0 {%k1}
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm9, %zmm0, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm6, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm6, %ymm14, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm30, %xmm11
; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm8, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm24 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm27 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa32 %zmm0, %zmm6 {%k1}
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm12[2,1,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm0, %zmm6, %zmm26
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm7, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpmovzxwd {{.*#+}} xmm28 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsp), %xmm12 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,0,2,1,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq %xmm12, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm29[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm24, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm4, %zmm31, %zmm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,8,3,4,9,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm21, %ymm4, %ymm30
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm22, %ymm4, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm27, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpermt2d %ymm28, %ymm4, %ymm6
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm30, %zmm4 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm4 = zmm30[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm14, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm4, %zmm14, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm23, %zmm4 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # zmm4 = zmm23[0,1,2,3],mem[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm13, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm4, %zmm14, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm19[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm14, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm6[0,1,2,3],zmm26[0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm12, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm14, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm27, %zmm8
+; AVX512F-ONLY-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm9, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm27, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm7, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm21, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm27, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm18[0,1,2,3]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm7, %zmm6
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm17, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 448(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 576(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm15, 512(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 704(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, 640(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $1272, %rsp # imm = 0x4F8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm20[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm7, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 448(%rax)
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm28[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm7, %zmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 640(%rax)
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm17[0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm7, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm16, %zmm0, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, (%rax)
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm26, %zmm0, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm25, %zmm1, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 320(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 384(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 576(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 512(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm0 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 704(%rax)
+; AVX512F-ONLY-FAST-NEXT: addq $1256, %rsp # imm = 0x4E8
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
; AVX512DQ-SLOW-LABEL: store_i16_stride6_vf64:
; AVX512DQ-SLOW: # %bb.0:
-; AVX512DQ-SLOW-NEXT: subq $392, %rsp # imm = 0x188
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512DQ-SLOW-NEXT: subq $936, %rsp # imm = 0x3A8
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %xmm0
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdx), %xmm2
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[8],ymm3[8],ymm4[9],ymm3[9],ymm4[10],ymm3[10],ymm4[11],ymm3[11]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm4, %ymm19
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [1,0,2,2,1,0,2,2,16,17,16,17,16,17,16,17]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm2, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, (%rsp) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm4[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm2
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %xmm2
; AVX512DQ-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm3, %xmm12
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm0[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm0[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %xmm9
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdx), %xmm8
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdx), %ymm3
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %xmm1
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %ymm7
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %ymm10
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm11 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm13
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm11, %zmm6, %zmm13 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %ymm5
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm5[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm13, %ymm6
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1],ymm4[2],ymm6[3,4],ymm4[5],ymm6[6,7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm6
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %xmm4
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm12, %xmm4, %xmm11
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm12, %xmm16
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm13[0],ymm11[1],ymm13[2,3],ymm11[4],ymm13[5,6],ymm11[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm11[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %xmm11
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm11[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm11, %xmm14
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm6[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %ymm6
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm11[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm7[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm10[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm11 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm7 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm10 = 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-SLOW-NEXT: vpsrldq {{.*#+}} ymm13 = 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-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm13[0],ymm10[0],ymm13[1],ymm10[1],ymm13[2],ymm10[2],ymm13[3],ymm10[3],ymm13[8],ymm10[8],ymm13[9],ymm10[9],ymm13[10],ymm10[10],ymm13[11],ymm10[11]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm2 = 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-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[1,2,3,3,5,6,7,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm11, %zmm3
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm2, %zmm10, %zmm3 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm7 = <u,u,u,u,u,u,u,u,14,15,14,15,14,15,14,15,28,29,26,27,26,27,30,31,30,31,30,31,30,31,30,31>
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm7, %ymm5, %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm7, %ymm15
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm3, %ymm7
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6],ymm2[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rcx), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdx), %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm2, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm5[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %xmm7
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm5[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm2, %zmm1 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm1, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm2
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1,2],ymm5[3],ymm3[4,5],ymm5[6],ymm3[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm3[0,1,2,3],zmm2[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %ymm12
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %ymm10
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm2 = 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-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm2[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rcx), %ymm7
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm11 = 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-SLOW-NEXT: vmovdqa 64(%rdx), %ymm3
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm13 = 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-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm11 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm13 = ymm3[4],ymm7[4],ymm3[5],ymm7[5],ymm3[6],ymm7[6],ymm3[7],ymm7[7],ymm3[12],ymm7[12],ymm3[13],ymm7[13],ymm3[14],ymm7[14],ymm3[15],ymm7[15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[1,2,3,3,5,6,7,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm6
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm13, %zmm11, %zmm6 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %ymm5
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm5, %ymm11
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm15, %ymm31
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm13
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm13[0],ymm11[1],ymm13[2,3],ymm11[4],ymm13[5,6],ymm11[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm6 = ymm13[0],ymm6[1,2],ymm13[3],ymm6[4,5],ymm13[6],ymm6[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm6[0,1,2,3],zmm11[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r9), %ymm6
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm11[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm6[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm11[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm11[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm1, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm4
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm1 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm2 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm31
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm19
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm0, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm3[0,1,2,1]
; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm15[0,1,2,1]
; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,7,6,5]
; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,0,2,2]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm8 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: movw $9362, %ax # imm = 0x2492
-; AVX512DQ-SLOW-NEXT: kmovw %eax, %k2
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm0
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm0 {%k2}
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm15[0],xmm3[0],xmm15[1],xmm3[1],xmm15[2],xmm3[2],xmm15[3],xmm3[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm18
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm3
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm3[2,1,3,3,4,5,6,7]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm8
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm8[1,2],ymm1[3],ymm8[4,5],ymm1[6],ymm8[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm4, %ymm4
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm14[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm14, %xmm17
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm0, %ymm22
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm2[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm8[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rcx), %xmm8
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdx), %xmm9
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[1,0,2,2]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm0 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm15 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm0, %ymm0
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm11
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm0, %zmm13, %zmm11 {%k2}
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %xmm4
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm4[2,1,3,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm11, %ymm13
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm13[1,2],ymm0[3],ymm13[4,5],ymm0[6],ymm13[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm13 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm13, %ymm13
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],ymm13[2],ymm11[3,4],ymm13[5],ymm11[6,7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm11[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0],ymm7[0],ymm3[1],ymm7[1],ymm3[2],ymm7[2],ymm3[3],ymm7[3],ymm3[8],ymm7[8],ymm3[9],ymm7[9],ymm3[10],ymm7[10],ymm3[11],ymm7[11]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm2 = 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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm2, %zmm1, %zmm0 {%k1}
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm16, %xmm14
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm14, %xmm4, %xmm2
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6],ymm2[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm30 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r9), %xmm4
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm0[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm0[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm8
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,2,3,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm9
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %ymm10
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm0[1,0,2,2,5,4,6,6]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm5
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm0[1,1,1,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm0
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm15 = 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]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm15, %zmm12, %zmm3 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm2[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm3, %ymm15
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm12 = ymm15[0,1],ymm12[2],ymm15[3,4],ymm12[5],ymm15[6,7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm12
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm15
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm14, %xmm15, %xmm14
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0],ymm14[1],ymm3[2,3],ymm14[4],ymm3[5,6],ymm14[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm3[0,1,2,3],zmm12[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm0[2,1,2,3,6,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm12[0],ymm3[0],ymm12[1],ymm3[1],ymm12[2],ymm3[2],ymm12[3],ymm3[3],ymm12[8],ymm3[8],ymm12[9],ymm3[9],ymm12[10],ymm3[10],ymm12[11],ymm3[11]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm0 = ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm1 = ymm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm9[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm12[0],ymm1[0],ymm12[1],ymm1[1],ymm12[2],ymm1[2],ymm12[3],ymm1[3],ymm12[8],ymm1[8],ymm12[9],ymm1[9],ymm12[10],ymm1[10],ymm12[11],ymm1[11]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm9 = 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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[1,2,3,3,5,6,7,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm9, %zmm1, %zmm0 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm31, %ymm12
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm12, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm26
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm2, %ymm2
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm3, %ymm28
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %ymm4
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 96(%rdx), %ymm11
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm4, %ymm21
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512DQ-SLOW-NEXT: kmovw %eax, %k1
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %ymm9
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm13, %ymm9, %ymm1
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm3
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm31 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm18, %ymm3
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm18[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm19, %ymm2
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm19[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
-; AVX512DQ-SLOW-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]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm3, %ymm22
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm24
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
-; AVX512DQ-SLOW-NEXT: vmovdqu (%rsp), %ymm10 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} ymm2 = 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-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 64(%rcx), %ymm12
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 64(%rdx), %ymm10
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpunpckhwd {{.*#+}} ymm3 = ymm9[4],ymm10[4],ymm9[5],ymm10[5],ymm9[6],ymm10[6],ymm9[7],ymm10[7],ymm9[12],ymm10[12],ymm9[13],ymm10[13],ymm9[14],ymm10[14],ymm9[15],ymm10[15]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm12, %ymm3, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm1
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm27, %zmm9
-; AVX512DQ-SLOW-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]
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm9 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm10
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm24, %zmm12, %zmm10
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm11
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm17, %xmm2
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm2[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm25
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm3[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm3 = xmm8[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm7
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,0,2,2]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k2}
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm14
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[3,3,3,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm7
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 32(%rdx), %ymm6
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[1,2,3,3,5,6,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm1
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm15[2,1,3,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0],ymm2[1,2],ymm3[3],ymm2[4,5],ymm3[6],ymm2[7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm6
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm3 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm3
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm6[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm1[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm4[0,1,2,1]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,7,6,5]
-; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm14[4],xmm8[4],xmm14[5],xmm8[5],xmm14[6],xmm8[6],xmm14[7],xmm8[7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm3[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm3[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm3, %ymm25
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0],ymm0[1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm5[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm0[2,1,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm0[3,3,3,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[1,2,3,3,5,6,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm17, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm0 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm1
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm1, %ymm8
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm13
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0],ymm8[1],ymm13[2,3],ymm8[4],ymm13[5,6],ymm8[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm0 = ymm13[0],ymm0[1,2],ymm13[3],ymm0[4,5],ymm13[6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm0[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm8
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm8[0,0,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm0, %ymm30
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm31, %xmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm19, %xmm13
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm13[4],xmm0[5],xmm13[5],xmm0[6],xmm13[6],xmm0[7],xmm13[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm8[0,2,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm13[0,0,2,1]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm4 = <17,18,17,18,u,u,19,19,5,4,2,2,5,4,6,6>
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm0, %zmm4, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm18, %xmm0
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %ymm15
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm15[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm13[2,2,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm15, %ymm16
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm5[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm5[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,1,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm0, %zmm2 {%k1}
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm26, %xmm0
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6],ymm0[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r9), %ymm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm0, %ymm15
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm5[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm2[0,1,2,3],zmm3[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm21, %ymm0
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermt2d %zmm2, %zmm4, %zmm11
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd (%rsp), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm28, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm27, %ymm3
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm11 {%k1}
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm9[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm11, %ymm3
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm9 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermt2d %zmm2, %zmm4, %zmm13
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm2 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm2 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,1,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm22, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm24, %ymm3
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm10 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermt2d %zmm10, %zmm4, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm4 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,1,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm3, %zmm2, %zmm13 {%k1}
+; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm6, %zmm4, %zmm5 {%k1}
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm3 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm4 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm7
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0],ymm7[1],ymm11[2,3],ymm7[4],ymm11[5,6],ymm7[7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm15[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm15, %ymm6
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm7[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm20 = zmm10[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm13, %ymm10
+; AVX512DQ-SLOW-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm11 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm11 = ymm10[0,1],ymm11[2],ymm10[3,4],ymm11[5],ymm10[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm12
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0],ymm12[1],ymm13[2,3],ymm12[4],ymm13[5,6],ymm12[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm10[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm13[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm12[0,1,2,3],zmm11[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm5, %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm15 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpsrldq {{.*#+}} xmm13 = xmm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,0,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm10[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm9
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0],ymm1[1],ymm5[2,3],ymm1[4],ymm5[5,6],ymm1[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm9[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm12
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm12[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,2,1,4,5,6,7]
; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm13, %ymm13
-; AVX512DQ-SLOW-NEXT: vinserti32x8 $1, %ymm13, %zmm8, %zmm6 {%k2}
-; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm6, %ymm8
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm4[2,1,3,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0],ymm8[1,2],ymm13[3],ymm8[4,5],ymm13[6],ymm8[7]
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm13[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,2,2,1,4,5,6,7,8,10,10,9,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-SLOW-NEXT: vpmovzxwd {{.*#+}} xmm24 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm24, %ymm5
-; AVX512DQ-SLOW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4],ymm5[5],ymm6[6,7]
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm6, %ymm6
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm8
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm8[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm8 = xmm4[0,0,2,1,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq %xmm8, %ymm8
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,4,4,4]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm30, %zmm12, %zmm11
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm16, %zmm12, %zmm2
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12, %zmm12 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm12 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm29, %zmm16
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm16 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm15, %zmm3
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm31, %zmm14, %zmm3
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm13, %zmm1
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm20, %zmm14, %zmm1
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm22, %zmm13
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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-SLOW-NEXT: vpternlogq $184, %zmm21, %zmm14, %zmm13
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm15
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm28, %zmm14, %zmm15
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm14, %zmm6
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm8, %zmm0
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm14, %zmm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm15 = mem[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[0,2,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm16, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm7 = mem[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm13 = mem[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,2,1,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm4 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm25
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm25 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm23, %zmm6
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm17, %zmm16
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm16 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm24, %zmm14
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm18, %zmm14 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm29, %zmm18, %zmm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm19, %zmm9, %zmm8
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm20, %zmm9, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm7, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm22, %zmm9, %zmm7
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm13, %zmm10
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm10
; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, (%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, 128(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, 320(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm11, 448(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm15, 384(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, 576(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, 256(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, 448(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, 640(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm25, (%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, 384(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 576(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm14, 320(%rax)
; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm16, 512(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, 704(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, 640(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, 64(%rax)
-; AVX512DQ-SLOW-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, 704(%rax)
+; AVX512DQ-SLOW-NEXT: addq $936, %rsp # imm = 0x3A8
; AVX512DQ-SLOW-NEXT: vzeroupper
; AVX512DQ-SLOW-NEXT: retq
;
; AVX512DQ-FAST-LABEL: store_i16_stride6_vf64:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $888, %rsp # imm = 0x378
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm11
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512DQ-FAST-NEXT: subq $1224, %rsp # imm = 0x4C8
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[8],ymm4[8],ymm0[9],ymm4[9],ymm0[10],ymm4[10],ymm0[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm19
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm0
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm7
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm6
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm6, %ymm30
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm31
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm8
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm8, %ymm28
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm17
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm16
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [1,1,1,1,10,10,10,11]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm18, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
+; AVX512DQ-FAST-NEXT: vpermd %zmm0, %zmm20, %zmm22
+; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
+; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm3, %zmm22 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm25 = [8,9,20,11,12,21,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm25, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm19 = [0,9,2,3,8,5,6,11]
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm19, %ymm22
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = <0,u,0,1,u,10,10,u>
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm5
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa 96(%rsi), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm14
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm21
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
+; AVX512DQ-FAST-NEXT: vpermd %zmm6, %zmm20, %zmm26
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm4, %zmm26 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm26, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm10
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm10[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm25, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = 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-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm13, %xmm17
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm24
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm19, %ymm26
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm11
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm4 = ymm11[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm23, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm3[4],xmm11[5],xmm3[5],xmm11[6],xmm3[6],xmm11[7],xmm3[7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm4, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm13
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm12
+; AVX512DQ-FAST-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-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm18, %zmm6
+; AVX512DQ-FAST-NEXT: vpermd %zmm17, %zmm20, %zmm17
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm6, %zmm17 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm17, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm6
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm6[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm7, %zmm25, %zmm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm15
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm15[0],ymm4[0],ymm15[1],ymm4[1],ymm15[2],ymm4[2],ymm15[3],ymm4[3],ymm15[8],ymm4[8],ymm15[9],ymm4[9],ymm15[10],ymm4[10],ymm15[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm10[4],xmm14[4],xmm10[5],xmm14[5],xmm10[6],xmm14[6],xmm10[7],xmm14[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm10, %xmm23
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm12
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <u,u,u,u,4,5,10,11,u,u,u,u,u,u,u,u,24,25,22,23,20,21,26,27,u,u,u,u,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm12, %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm14
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm14, %ymm10
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[8],ymm6[8],ymm10[9],ymm6[9],ymm10[10],ymm6[10],ymm10[11],ymm6[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm10 = ymm14[4],ymm12[4],ymm14[5],ymm12[5],ymm14[6],ymm12[6],ymm14[7],ymm12[7],ymm14[12],ymm12[12],ymm14[13],ymm12[13],ymm14[14],ymm12[14],ymm14[15],ymm12[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = [5,6,5,6,5,6,7,7]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm12 = ymm8[4],ymm5[4],ymm8[5],ymm5[5],ymm8[6],ymm5[6],ymm8[7],ymm5[7],ymm8[12],ymm5[12],ymm8[13],ymm5[13],ymm8[14],ymm5[14],ymm8[15],ymm5[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm12, %ymm18, %ymm12
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm5 = ymm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm5[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm21
-; AVX512DQ-FAST-NEXT: movw $18724, %ax # imm = 0x4924
-; AVX512DQ-FAST-NEXT: kmovw %eax, %k1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm12, %zmm5, %zmm21 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [8,21,10,11,20,13,14,23]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm21, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm13, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa %ymm13, %ymm12
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm20, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [12,1,2,13,4,5,14,7]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm5 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm16, %ymm21
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm13, %ymm6, %ymm5
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm8, %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm10, %ymm6
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = 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]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm8 = ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm18, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm7
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm7, %ymm19, %ymm17
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %ymm7
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm7[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-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-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm18, %zmm0
+; AVX512DQ-FAST-NEXT: vpermd %zmm16, %zmm20, %zmm20
+; AVX512DQ-FAST-NEXT: vmovdqa32 %zmm0, %zmm20 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm0, %zmm20, %zmm25
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm19, %ymm20
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm0, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm23, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm15, %ymm9, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm5, %ymm1
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[8],ymm0[8],ymm1[9],ymm0[9],ymm1[10],ymm0[10],ymm1[11],ymm0[11]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm29 = [2,1,2,3,11,11,11,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm29, %zmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm27 = [5,6,5,6,5,6,7,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm30, %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512DQ-FAST-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-FAST-NEXT: vpermd %ymm0, %ymm27, %ymm0
; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm1 = ymm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm1[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm2 = ymm2[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm2[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11]
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm27
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm1, %zmm27 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, %zmm29
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm20, %zmm29
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm2[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm16, %ymm27
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm31
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm13, %ymm30
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm1, %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm6, %xmm5
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm2[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm8 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [8,21,10,11,20,13,14,23]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm0
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm2, %ymm10, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm2, %ymm9
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm23, %zmm0
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm10[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm30 = [12,1,2,13,4,5,14,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm30, %ymm8
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm0[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,u,u,u,u,u,u,8,9,10,11,12,13,14,15,24,25,28,29,28,29,26,27,24,25,26,27,28,29,30,31>
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm1, %ymm8
+; AVX512DQ-FAST-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-FAST-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm2, %ymm8
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm1[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm9[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm9, %ymm9
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm25
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm2, %ymm11
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = <2,2,u,3,10,u,10,11>
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm12, %ymm2
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[1],ymm0[1],ymm2[2],ymm0[2],ymm2[3],ymm0[3],ymm2[8],ymm0[8],ymm2[9],ymm0[9],ymm2[10],ymm0[10],ymm2[11],ymm0[11]
+; AVX512DQ-FAST-NEXT: 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-FAST-NEXT: vpermt2q %zmm2, %zmm29, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm28, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm2 = ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[12],ymm3[12],ymm4[13],ymm3[13],ymm4[14],ymm3[14],ymm4[15],ymm3[15]
+; AVX512DQ-FAST-NEXT: vpermd %ymm2, %ymm27, %ymm2
+; AVX512DQ-FAST-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-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} 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-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm2, %zmm3, %zmm0 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vpshufb %ymm9, %ymm6, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm9, %ymm10
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm23, %zmm2
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm3 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm3, %ymm30, %ymm0
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm7, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm7, %ymm2
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm31, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,6,7,8,9,6,7,4,5,10,11]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm21, %xmm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm4, %xmm0
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm14, %xmm3
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [0,0,2,1,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm19, %zmm3
+; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [1,0,2,2,1,0,2,2]
+; AVX512DQ-FAST-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm4, %ymm0, %ymm4
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
; AVX512DQ-FAST-NEXT: movw $9362, %ax # imm = 0x2492
; AVX512DQ-FAST-NEXT: kmovw %eax, %k2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm9, %zmm8, %zmm25 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm22 = [16,9,10,17,12,13,18,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm25, %zmm26
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm22, %zmm26
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm1, %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm3[0],xmm11[1],xmm3[1],xmm11[2],xmm3[2],xmm11[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm2, %ymm8
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm3 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm9 = xmm11[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm3[0],xmm9[1],xmm3[1],xmm9[2],xmm3[2],xmm9[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm3, %ymm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm11
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm3, %zmm8, %zmm11 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm11, %zmm28
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm3, %zmm22, %zmm28
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm6
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm6, %ymm5
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm4, %zmm3 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [16,9,10,17,12,13,18,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm6
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm28, %zmm4
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm24[0],zero,xmm24[1],zero,xmm24[2],zero,xmm24[3],zero
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm16 = [0,1,8,3,4,9,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm16, %ymm3
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm4[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm4, %xmm5, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm4, %xmm8
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = <u,0,0,u,8,8,u,9>
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm18, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm5, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm4
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm5 = xmm5[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm6 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm3, %ymm0, %ymm3
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm3, %zmm4 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm28, %zmm3
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm5 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm16, %ymm4
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm4[0,1,2,3],zmm3[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm4, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm8, %xmm21
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm18, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm8, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm6, %ymm5
; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm5[0],ymm3[0],ymm5[1],ymm3[1],ymm5[2],ymm3[2],ymm5[3],ymm3[3],ymm5[8],ymm3[8],ymm5[9],ymm3[9],ymm5[10],ymm3[10],ymm5[11],ymm3[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm5 = ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[12],ymm1[12],ymm6[13],ymm1[13],ymm6[14],ymm1[14],ymm6[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm6 = ymm15[4],ymm4[4],ymm15[5],ymm4[5],ymm15[6],ymm4[6],ymm15[7],ymm4[7],ymm15[12],ymm4[12],ymm15[13],ymm4[13],ymm15[14],ymm4[14],ymm15[15],ymm4[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm6, %ymm18, %ymm6
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm4 = ymm4[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm4[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm0 = ymm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm15[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[8],ymm4[8],ymm0[9],ymm4[9],ymm0[10],ymm4[10],ymm0[11],ymm4[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm0, %zmm4 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, %zmm17
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa %ymm12, %ymm10
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm20, %zmm17
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm0 = ymm1[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm16, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm13
-; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm0[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm30, %ymm15
-; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm0[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm8
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm8, %ymm5
-; AVX512DQ-FAST-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]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,2,3]
+; AVX512DQ-FAST-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-FAST-NEXT: vpermt2q %zmm5, %zmm29, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX512DQ-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm19, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} ymm7 = ymm3[4],ymm1[4],ymm3[5],ymm1[5],ymm3[6],ymm1[6],ymm3[7],ymm1[7],ymm3[12],ymm1[12],ymm3[13],ymm1[13],ymm3[14],ymm1[14],ymm3[15],ymm1[15]
-; AVX512DQ-FAST-NEXT: vpermd %ymm7, %ymm18, %ymm7
-; AVX512DQ-FAST-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-FAST-NEXT: vpsrldq {{.*#+}} ymm12 = ymm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm3[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} 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]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm7, %zmm9, %zmm1 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm5
-; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm7
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm7, %zmm1, %zmm20
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm7 = ymm5[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm7, %ymm16, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm0, %ymm7
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm7[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm7
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm7[2,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm14, %xmm13
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm9, %xmm12
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm7, %xmm14
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm13, %xmm10
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm12[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm23, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm15, %ymm2, %ymm15
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm12 = xmm3[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm13 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm12, %ymm12
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm14, %zmm16
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm12, %zmm15, %zmm16 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, %zmm23
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm12
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm12[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm13, %zmm22, %zmm23
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm14
-; AVX512DQ-FAST-NEXT: vpshufb %xmm10, %xmm13, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufb %xmm10, %xmm14, %xmm15
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm0[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vpermd %ymm15, %ymm2, %ymm15
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm10 = xmm10[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm2, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm0, %zmm15, %zmm2 {%k2}
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm0
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[2,1,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm3, %zmm2, %zmm22
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} ymm6 = ymm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,ymm6[22,23,24,25,26,27,28,29,30,31],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: 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-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm8 = xmm8[4],mem[4],xmm8[5],mem[5],xmm8[6],mem[6],xmm8[7],mem[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[1,1,1,1]
+; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm27, %ymm5
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm5, %zmm6, %zmm3 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm10, %ymm14
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm8, %ymm5
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm23, %zmm6
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm5 = ymm8[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm5, %ymm30, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm5
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm10 = ymm10[0],mem[0],ymm10[1],mem[1],ymm10[2],mem[2],ymm10[3],mem[3],ymm10[8],mem[8],ymm10[9],mem[9],ymm10[10],mem[10],ymm10[11],mem[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm10, %ymm8
+; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vpermt2q %zmm8, %zmm29, %zmm5
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm29 = zmm3[0,1,2,3],zmm6[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm13, %ymm3, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm11, %ymm15
+; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm31, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512DQ-FAST-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-FAST-NEXT: vpermd %ymm8, %ymm27, %ymm8
+; AVX512DQ-FAST-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-FAST-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-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm8, %zmm10, %zmm5 {%k1}
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm6, %ymm8
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm8, %zmm5, %zmm23
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm8 = ymm6[2,1,3,3,4,5,6,7,10,9,11,11,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm8, %ymm30, %ymm5
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm6, %ymm10
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm6, %ymm8
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm31, %zmm8
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm10
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm13 = xmm13[4],mem[4],xmm13[5],mem[5],xmm13[6],mem[6],xmm13[7],mem[7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm14 = ymm14[0],mem[0],ymm14[1],mem[1],ymm14[2],mem[2],ymm14[3],mem[3],ymm14[8],mem[8],ymm14[9],mem[9],ymm14[10],mem[10],ymm14[11],mem[11]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm14[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm21[0,1,2,3],zmm24[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm21, %zmm24 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm14, %zmm21, %zmm24
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm27[0,1,2,3],zmm29[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm27 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm27, %zmm27 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm14, %zmm21, %zmm27
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm17[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm31, %zmm29
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm4, %zmm21, %zmm29
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm20[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm19, %zmm30
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm21, %zmm30
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm13, %xmm11
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm19, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm17 = [0,1,8,3,4,9,6,7]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm17, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm17, %ymm11
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm12[0],zero,xmm12[1],zero,xmm12[2],zero,xmm12[3],zero
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm17, %ymm16
-; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm17, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = <1,2,1,2,u,u,3,3,13,12,10,10,13,12,14,14>
-; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm20 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm6, %zmm3, %zmm20 {%k1}
-; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm19 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm10, %zmm8, %zmm19 {%k1}
-; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm8 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm15, %zmm13, %zmm8 {%k1}
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm18 = [8,9,20,11,12,21,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, %zmm21
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm1, %zmm18, %zmm21
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = [12,13,10,11,10,11,14,15,14,15,14,15,14,15,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [0,9,2,3,8,5,6,11]
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm15, %ymm20
-; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm0 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, %zmm10
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm0, %zmm18, %zmm10
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm25[0,1,2,3],zmm26[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm14, %xmm1
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm1, %ymm15, %ymm19
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm14 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm14 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[8],mem[8],ymm1[9],mem[9],ymm1[10],mem[10],ymm1[11],mem[11]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm7
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm7[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm1, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,4,5,4,5,6,7,8,9,8,9,8,9,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm7, %xmm3
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm31
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm25, %zmm31
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm28 = zmm11[0,1,2,3],zmm28[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm0 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[1,1,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm14, %zmm5, %zmm0 {%k1}
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm14
-; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm5 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, %zmm17
-; AVX512DQ-FAST-NEXT: vpermt2d %zmm5, %zmm18, %zmm17
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm14[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm5, %ymm5
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm14, %xmm11
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm5, %zmm5
-; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm4, %xmm11
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm11, %ymm15, %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm11
-; AVX512DQ-FAST-NEXT: vpshuflw $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,1,1,3,4,5,6,7,8,9,9,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm4, %zmm0, %zmm18
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm4, %ymm26
-; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm12, %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm12
-; AVX512DQ-FAST-NEXT: vpermt2d %ymm6, %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm11, %xmm6
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm12, %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [8,9,12,13,12,13,10,11,8,9,10,11,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm7, %xmm7
-; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm14, %xmm14
-; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm11, %xmm11
-; AVX512DQ-FAST-NEXT: vpshufb %xmm15, %xmm12, %xmm15
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,0,2,1,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm12, %ymm12
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm3 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshuflw $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm13 = mem[0,0,2,1,4,5,6,7,8,8,10,9,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm28, %zmm25, %zmm5
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm16[0,1,2,3],zmm23[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm26, %zmm6
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm16, %zmm25, %zmm6
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm22[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm12, %zmm9
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm25, %zmm9
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm20[0,1,2,3],zmm21[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm1
-; AVX512DQ-FAST-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]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm12, %zmm1
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm19[0,1,2,3],zmm10[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm12, %zmm4
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm8[0,1,2,3],zmm17[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm14, %zmm3
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm12, %zmm3
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm18[0,1,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm12, %zmm2
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm0, %ymm11
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm13 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm14 = xmm14[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm13, %ymm13
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm13, %zmm11, %zmm10 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, %zmm11
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsp), %xmm15 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm15[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2d %zmm13, %zmm28, %zmm11
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm14, %xmm13
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm9
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm6[0],xmm14[0],xmm6[1],xmm14[1],xmm6[2],xmm14[2],xmm6[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm19, %zmm13
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm9 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm6[0],xmm15[0],xmm6[1],xmm15[1],xmm6[2],xmm15[2],xmm6[3],xmm15[3]
+; AVX512DQ-FAST-NEXT: vpermd %ymm14, %ymm0, %ymm0
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm14 = xmm15[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpsrldq {{.*#+}} xmm15 = xmm6[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vpbroadcastq %xmm14, %ymm14
+; AVX512DQ-FAST-NEXT: vinserti32x8 $1, %ymm14, %zmm0, %zmm13 {%k2}
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[2,1,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm0, %zmm13, %zmm28
+; AVX512DQ-FAST-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm9, %ymm16, %ymm10
+; AVX512DQ-FAST-NEXT: vpermt2d %ymm0, %ymm16, %ymm13
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm5[0,1,2,3],zmm23[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm10[0,1,2,3],zmm11[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm21, %xmm11
+; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm9
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm13[0,1,2,3],zmm28[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm11
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm6[0,0,2,1,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm11, %zmm18, %zmm13
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $184, %zmm9, %zmm11, %zmm13
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 448(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 640(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, (%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm11, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 192(%rax)
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm29, %zmm5, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 320(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm20[0,1,2,3],zmm25[0,1,2,3]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # zmm0 = zmm17[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 448(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm12, %zmm11, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 384(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 576(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 512(%rax)
+; AVX512DQ-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 704(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # zmm0 = zmm26[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 640(%rax)
+; AVX512DQ-FAST-NEXT: vshufi64x2 $68, {{[-0-9]+}}(%r{{[sb]}}p), %zmm22, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # zmm0 = zmm22[0,1,2,3],mem[0,1,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm1
; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 64(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm30, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm29, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 384(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm31, 576(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, 512(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, 704(%rax)
-; AVX512DQ-FAST-NEXT: addq $888, %rsp # imm = 0x378
+; AVX512DQ-FAST-NEXT: addq $1224, %rsp # imm = 0x4C8
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index 64be8677d5a49dc..473ac8a546f9048 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
@@ -2315,285 +2315,539 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
-; AVX512F-SLOW-LABEL: store_i16_stride7_vf16:
-; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa (%r10), %ymm1
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u],zero,zero,ymm4[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[16,17,u,u,u,u,u,u,u,u]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,u,u,u,14,15],zero,zero,ymm3[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm3[u,u,u,u,u,u,u,u]
-; AVX512F-SLOW-NEXT: vpor %ymm7, %ymm8, %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm8
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm9
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm7
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = zero,zero,zero,zero,ymm6[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm6[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[12,13,14,15],zero,zero,ymm5[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm5[u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512F-SLOW-NEXT: vpor %ymm10, %ymm11, %ymm12
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm10
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm11
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm15
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm15
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = 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]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm0[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-SLOW-NEXT: vpor %ymm7, %ymm12, %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm12
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm13
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm14, %xmm18
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpbroadcastd 8(%r10), %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm16 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpandnq %ymm7, %ymm16, %ymm7
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm14 = ymm1[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[16,17,u,u]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm7, %zmm7
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512F-SLOW-NEXT: vprold $16, %xmm11, %xmm14
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm10[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0,1],xmm14[2],xmm15[3,4],xmm14[5],xmm15[6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0],xmm8[1],xmm9[2,3],xmm8[4],xmm9[5,6],xmm8[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm18, %xmm1
-; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm9, %xmm9
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm8[0,0,1,1,4,4,5,5]
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm8 = xmm0[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,1,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512F-SLOW-NEXT: vpbroadcastd (%r10), %ymm8
-; AVX512F-SLOW-NEXT: vpbroadcastd 4(%r10), %ymm10
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7,8,9],ymm9[10],ymm0[11,12],ymm9[13],ymm0[14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm10 = ymm6[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3],ymm10[4],ymm9[5,6,7,8],ymm10[9],ymm9[10,11],ymm10[12],ymm9[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4],ymm9[5],ymm10[6,7,8,9],ymm9[10],ymm10[11,12],ymm9[13],ymm10[14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm5[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm6[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm17[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1,2],ymm0[3],ymm10[4,5],ymm0[6],ymm10[7,8,9,10],ymm0[11],ymm10[12,13],ymm0[14],ymm10[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512F-SLOW-NEXT: vprold $16, %ymm2, %ymm10
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm17[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm16[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm16[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm10
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm5[3],ymm0[4,5],ymm5[6],ymm0[7,8,9,10],ymm5[11],ymm0[12,13],ymm5[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3],ymm4[4],ymm3[5,6,7,8],ymm4[9],ymm3[10,11],ymm4[12],ymm3[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa %ymm1, 192(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm10, 128(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm8, (%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
-; AVX512F-SLOW-NEXT: vzeroupper
-; AVX512F-SLOW-NEXT: retq
+; AVX512F-ONLY-SLOW-LABEL: store_i16_stride7_vf16:
+; AVX512F-ONLY-SLOW: # %bb.0:
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rax), %ymm13
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u],zero,zero,ymm7[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[16,17,u,u,u,u,u,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,14,15],zero,zero,ymm6[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm6[u,u,u,u,u,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vporq %ymm1, %ymm4, %ymm16
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm0, %xmm21
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,ymm9[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm9[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[12,13,14,15],zero,zero,ymm8[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm8[u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512F-ONLY-SLOW-NEXT: vporq %ymm1, %ymm4, %ymm17
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm4
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vporq %ymm5, %ymm10, %ymm19
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm5, %ymm10, %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm13[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm13[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm13[16,17,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm10[2],ymm12[3,4],ymm10[5],ymm12[6,7,8,9],ymm10[10],ymm12[11,12],ymm10[13],ymm12[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm8[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm20
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm4, %xmm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm1[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm11 = xmm12[0,1],xmm11[2],xmm12[3,4],xmm11[5],xmm12[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = [0,16,0,1,17,17,2,0,0,16,0,1,17,17,2,0]
+; AVX512F-ONLY-SLOW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm12, %zmm11, %zmm18
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm8[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm11[0],ymm12[1],ymm11[2,3],ymm12[4],ymm11[5,6,7,8],ymm12[9],ymm11[10,11],ymm12[12],ymm11[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0],xmm14[1],xmm15[2,3],xmm14[4],xmm15[5,6],xmm14[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1,2],ymm0[3],ymm14[4,5],ymm0[6],ymm14[7,8,9,10],ymm0[11],ymm14[12,13],ymm0[14],ymm14[15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm0[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm11, %zmm0, %zmm15
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm3, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm2[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3],ymm8[4,5],ymm9[6],ymm8[7,8,9,10],ymm9[11],ymm8[12,13],ymm9[14],ymm8[15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3],ymm7[4],ymm6[5,6,7,8],ymm7[9],ymm6[10,11],ymm7[12],ymm6[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm21, %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm10[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm22[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm12[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm14[2,2,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512F-ONLY-SLOW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermd %zmm13, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,3,3,3,6,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rcx
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm4, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm20[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm12, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm1[0,0,1,1,4,4,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd (%rax), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm13[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm0, 192(%rcx)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, (%rcx)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 128(%rcx)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, 64(%rcx)
+; AVX512F-ONLY-SLOW-NEXT: vzeroupper
+; AVX512F-ONLY-SLOW-NEXT: retq
;
-; AVX512F-FAST-LABEL: store_i16_stride7_vf16:
-; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm4
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm1
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,ymm5[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm3[12,13,14,15],zero,zero,ymm3[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm3[u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512F-FAST-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm8
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm9
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u],zero,zero,ymm4[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[16,17,u,u,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,14,15],zero,zero,ymm2[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm2[u,u,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpor %ymm7, %ymm10, %ymm7
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm10
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-FAST-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,2,3,3,4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm12, %zmm15
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm15
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm1[u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[16,17,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-FAST-NEXT: vpor %ymm6, %ymm7, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm12
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm13
-; AVX512F-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm7, %xmm7
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm0, %xmm18
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm16
-; AVX512F-FAST-NEXT: vpbroadcastd 8(%r10), %ymm6
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-FAST-NEXT: vpandn %ymm6, %ymm7, %ymm6
-; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm7[16,17,u,u]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm6, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm6
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm6
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,1,3]
-; AVX512F-FAST-NEXT: vprold $16, %xmm9, %xmm15
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm8[1,1,2,3]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm15[2],xmm0[3,4],xmm15[5],xmm0[6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm14, %zmm0
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512F-FAST-NEXT: vmovdqa64 %xmm18, %xmm9
-; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm8, %xmm8
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,1,1,4,4,5,5]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm0[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
-; AVX512F-FAST-NEXT: vpbroadcastd (%r10), %ymm9
-; AVX512F-FAST-NEXT: vpbroadcastd 4(%r10), %ymm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512F-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm2[0,1,1,3,4,5,5,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1],ymm0[2],ymm8[3,4],ymm0[5],ymm8[6,7,8,9],ymm0[10],ymm8[11,12],ymm0[13],ymm8[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2,3],ymm8[4],ymm10[5,6,7,8],ymm8[9],ymm10[10,11],ymm8[12],ymm10[13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm10[0,1],ymm8[2],ymm10[3,4],ymm8[5],ymm10[6,7,8,9],ymm8[10],ymm10[11,12],ymm8[13],ymm10[14,15]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,u,u,20,21,24,25,u,u,22,23,22,23]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm17[0,0,2,1,4,4,6,5]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,u,u,18,19,20,21,u,u,20,21]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1,2],ymm0[3],ymm10[4,5],ymm0[6],ymm10[7,8,9,10],ymm0[11],ymm10[12,13],ymm0[14],ymm10[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm10
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm17[1,2,2,3,5,6,6,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512F-FAST-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-FAST-NEXT: vpermd %zmm7, %zmm10, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm10
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,14,15,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1,2],ymm0[3],ymm3[4,5],ymm0[6],ymm3[7,8,9,10],ymm0[11],ymm3[12,13],ymm0[14],ymm3[15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,3]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u,u,u,u,u]
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm2
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <6,u,u,u,7,u,u,7>
-; AVX512F-FAST-NEXT: vpermd %ymm7, %ymm0, %ymm0
-; AVX512F-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
-; AVX512F-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512F-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa %ymm1, 192(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, 128(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm9, (%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 64(%rax)
-; AVX512F-FAST-NEXT: vzeroupper
-; AVX512F-FAST-NEXT: retq
+; AVX512F-ONLY-FAST-LABEL: store_i16_stride7_vf16:
+; AVX512F-ONLY-FAST: # %bb.0:
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = 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,zero,zero,zero,zero,zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[12,13,14,15],zero,zero,ymm5[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm5[u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm3, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u],zero,zero,ymm6[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[16,17,u,u,u,u,u,u,u,u]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm3, %ymm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm3, %ymm18
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm9[0,1],ymm3[2],ymm9[3,4],ymm3[5],ymm9[6,7,8,9],ymm3[10],ymm9[11,12],ymm3[13],ymm9[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = <2,u,3,2,u,10,10,11>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,u,u,20,21,24,25,u,u,22,23,22,23]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm10[0,1],ymm3[2],ymm10[3,4],ymm3[5],ymm10[6,7,8,9],ymm3[10],ymm10[11,12],ymm3[13],ymm10[14,15]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm2, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm13 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,u,u,18,19,20,21,u,u,20,21]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm13[0,1,2],ymm3[3],ymm13[4,5],ymm3[6],ymm13[7,8,9,10],ymm3[11],ymm13[12,13],ymm3[14],ymm13[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [2,2,3,3,10,9,11,10]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm12, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2],xmm3[3,4],xmm0[5],xmm3[6,7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,8,1,9,0,8,1,9]
+; AVX512F-ONLY-FAST-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm15
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm14[1],xmm8[2,3],xmm14[4],xmm8[5,6],xmm14[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm14
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm11, %xmm11
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm12
+; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 8(%rax), %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpandnq %ymm10, %ymm20, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [0,0,0,1,8,9,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm14, %zmm8, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm8[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[16,17,u,u]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,14,15,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3],ymm6[4],ymm4[5,6,7,8],ymm6[9],ymm4[10,11],ymm6[12],ymm4[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512F-ONLY-FAST-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm8, %zmm6, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <6,u,u,u,7,u,u,7>
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm7, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm12[0,0,1,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rcx
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm8, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm19[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd (%rax), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 4(%rax), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm1, 192(%rcx)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, (%rcx)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 128(%rcx)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 64(%rcx)
+; AVX512F-ONLY-FAST-NEXT: vzeroupper
+; AVX512F-ONLY-FAST-NEXT: retq
+;
+; AVX512DQ-SLOW-LABEL: store_i16_stride7_vf16:
+; AVX512DQ-SLOW: # %bb.0:
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm9
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm6
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rax), %ymm13
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u],zero,zero,ymm7[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm7[16,17,u,u,u,u,u,u,u,u]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm6[u,u,u,u,u,u,14,15],zero,zero,ymm6[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm6[u,u,u,u,u,u,u,u]
+; AVX512DQ-SLOW-NEXT: vporq %ymm1, %ymm4, %ymm16
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm0, %xmm21
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,ymm9[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm9[16,17,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[12,13,14,15],zero,zero,ymm8[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm8[u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512DQ-SLOW-NEXT: vporq %ymm1, %ymm4, %ymm17
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm4
+; AVX512DQ-SLOW-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-SLOW-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-SLOW-NEXT: vporq %ymm5, %ymm10, %ymm19
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm5
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm5, %ymm10, %ymm5
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm13[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm13[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm13[16,17,u,u]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = 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-SLOW-NEXT: vmovdqa64 %ymm0, %ymm22
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm7[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-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm10[2],ymm12[3,4],ymm10[5],ymm12[6,7,8,9],ymm10[10],ymm12[11,12],ymm10[13],ymm12[14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm8[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm11 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm10[2],ymm11[3,4],ymm10[5],ymm11[6,7,8,9],ymm10[10],ymm11[11,12],ymm10[13],ymm11[14,15]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm20
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm4, %xmm11
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm1[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm11 = xmm12[0,1],xmm11[2],xmm12[3,4],xmm11[5],xmm12[6,7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm18 = [0,16,0,1,17,17,2,0,0,16,0,1,17,17,2,0]
+; AVX512DQ-SLOW-NEXT: # zmm18 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm12, %zmm11, %zmm18
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm8[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm11[0],ymm12[1],ymm11[2,3],ymm12[4],ymm11[5,6,7,8],ymm12[9],ymm11[10,11],ymm12[12],ymm11[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm11
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0],xmm14[1],xmm15[2,3],xmm14[4],xmm15[5,6],xmm14[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm15
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1,2],ymm0[3],ymm14[4,5],ymm0[6],ymm14[7,8,9,10],ymm0[11],ymm14[12,13],ymm0[14],ymm14[15]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm0[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm11, %zmm0, %zmm15
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm3, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm2[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3],ymm8[4,5],ymm9[6],ymm8[7,8,9,10],ymm9[11],ymm8[12,13],ymm9[14],ymm8[15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3],ymm7[4],ymm6[5,6,7,8],ymm7[9],ymm6[10,11],ymm7[12],ymm6[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm21, %xmm7
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm10[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm22[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm12[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm14[2,2,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm3 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512DQ-SLOW-NEXT: # zmm3 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermd %zmm13, %zmm3, %zmm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,3,3,3,6,7,7,7]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm7, %zmm7
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rcx
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm9
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm9
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm4, %zmm4
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm4
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm20[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm7
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm12, %zmm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm3
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm1[0,0,1,1,4,4,5,5]
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm0
+; AVX512DQ-SLOW-NEXT: vpbroadcastd (%rax), %ymm1
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm4
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm6
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm13[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm0, 192(%rcx)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, (%rcx)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm3, 128(%rcx)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, 64(%rcx)
+; AVX512DQ-SLOW-NEXT: vzeroupper
+; AVX512DQ-SLOW-NEXT: retq
+;
+; AVX512DQ-FAST-LABEL: store_i16_stride7_vf16:
+; AVX512DQ-FAST: # %bb.0:
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm0 = 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,zero,zero,zero,zero,zero,zero
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm5[12,13,14,15],zero,zero,ymm5[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm5[u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512DQ-FAST-NEXT: vporq %ymm0, %ymm3, %ymm16
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm11
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u],zero,zero,ymm6[14,15,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm6[16,17,u,u,u,u,u,u,u,u]
+; AVX512DQ-FAST-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-FAST-NEXT: vporq %ymm0, %ymm3, %ymm17
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm14
+; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vporq %ymm0, %ymm3, %ymm18
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm9[0,1],ymm3[2],ymm9[3,4],ymm3[5],ymm9[6,7,8,9],ymm3[10],ymm9[11,12],ymm3[13],ymm9[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = <2,u,3,2,u,10,10,11>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm9
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm6[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-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,u,u,20,21,24,25,u,u,22,23,22,23]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm10[0,1],ymm3[2],ymm10[3,4],ymm3[5],ymm10[6,7,8,9],ymm3[10],ymm10[11,12],ymm3[13],ymm10[14,15]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm19
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm2, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm13 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,u,u,18,19,20,21,u,u,20,21]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm13[0,1,2],ymm3[3],ymm13[4,5],ymm3[6],ymm13[7,8,9,10],ymm3[11],ymm13[12,13],ymm3[14],ymm13[15]
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm13
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm12, %xmm0
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2],xmm3[3,4],xmm0[5],xmm3[6,7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm15 = [0,8,1,9,0,8,1,9]
+; AVX512DQ-FAST-NEXT: # zmm15 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm15
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm14[1],xmm8[2,3],xmm14[4],xmm8[5,6],xmm14[7]
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm14
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm11, %xmm11
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm11
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm8
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm12
+; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-FAST-NEXT: vpbroadcastd 8(%rax), %ymm10
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm20 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpandnq %ymm10, %ymm20, %ymm10
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm14, %zmm8, %zmm20
+; AVX512DQ-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm8[12,13,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[14,15,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm8[16,17,u,u]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm10
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,14,15,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,u,u,u,u,30,31,u,u,u,u,u,u]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3],ymm6[4],ymm4[5,6,7,8],ymm6[9],ymm4[10,11],ymm6[12],ymm4[13,14,15]
+; AVX512DQ-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512DQ-FAST-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FAST-NEXT: vpermd %zmm8, %zmm6, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = <6,u,u,u,7,u,u,7>
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm7, %ymm7
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm12[0,0,1,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,2,3]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rcx
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm2
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm8, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm10
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm19[2,2,2,3,6,6,6,7]
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm6
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm6
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm0
+; AVX512DQ-FAST-NEXT: vpbroadcastd (%rax), %ymm2
+; AVX512DQ-FAST-NEXT: vpbroadcastd 4(%rax), %ymm3
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm4
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm1
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm1, 192(%rcx)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, (%rcx)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 128(%rcx)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 64(%rcx)
+; AVX512DQ-FAST-NEXT: vzeroupper
+; AVX512DQ-FAST-NEXT: retq
;
; AVX512BW-LABEL: store_i16_stride7_vf16:
; AVX512BW: # %bb.0:
@@ -4996,1010 +5250,1243 @@ define void @store_i16_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
-; AVX512F-SLOW-LABEL: store_i16_stride7_vf32:
-; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: subq $824, %rsp # imm = 0x338
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm22
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm4, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm21
-; AVX512F-SLOW-NEXT: vpor %ymm1, %ymm3, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm6
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm6, %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm6, %ymm20
-; AVX512F-SLOW-NEXT: vpor %ymm3, %ymm4, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm6
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm6, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm8
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm9, %ymm8, %ymm7
-; AVX512F-SLOW-NEXT: vpor %ymm3, %ymm7, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm12
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm14
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa %ymm3, %ymm13
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm3, %ymm5
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512F-SLOW-NEXT: vpor %ymm1, %ymm5, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm3, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm1
-; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm1, %ymm2
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %ymm2
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm2, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm5
-; AVX512F-SLOW-NEXT: vpshufb %ymm9, %ymm5, %ymm4
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm4, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
-; AVX512F-SLOW-NEXT: vpermi2d %zmm4, %zmm0, %zmm7
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm8[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm6[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm10 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm11 = xmm9[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermi2d %zmm11, %zmm10, %zmm7
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm9[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
-; AVX512F-SLOW-NEXT: vpermi2d %zmm4, %zmm0, %zmm7
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vprold $16, %ymm6, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm8[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm8[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm0[3],ymm6[4,5],ymm0[6],ymm6[7,8,9,10],ymm0[11],ymm6[12,13],ymm0[14],ymm6[15]
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm7
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,2,3,3,10,9,11,10]
-; AVX512F-SLOW-NEXT: vpermi2q %zmm4, %zmm6, %zmm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm11
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm4[0,0,2,1]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rax), %ymm8
-; AVX512F-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpandn %ymm4, %ymm6, %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm8, %ymm10
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm4, %zmm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
-; AVX512F-SLOW-NEXT: vmovdqa %ymm13, %ymm0
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm13[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm17[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = 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]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm9, %ymm29
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm12[0,1],ymm10[2],ymm12[3,4],ymm10[5],ymm12[6,7,8,9],ymm10[10],ymm12[11,12],ymm10[13],ymm12[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm13, %ymm30
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rax), %ymm14
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,2,2,3]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpandn %ymm10, %ymm12, %ymm10
-; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm14, %ymm6
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm5[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm6[0,1,2],ymm10[3],ymm6[4,5],ymm10[6],ymm6[7,8,9,10],ymm10[11],ymm6[12,13],ymm10[14],ymm6[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm12, %ymm31
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = 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]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm10[2],ymm6[3,4],ymm10[5],ymm6[6,7,8,9],ymm10[10],ymm6[11,12],ymm10[13],ymm6[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0],ymm6[1],ymm1[2,3],ymm6[4],ymm1[5,6,7,8],ymm6[9],ymm1[10,11],ymm6[12],ymm1[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm13, %ymm27
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm6[0],ymm1[1],ymm6[2,3],ymm1[4],ymm6[5,6,7,8],ymm1[9],ymm6[10,11],ymm1[12],ymm6[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm13, %ymm19
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm0[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm1[0,1,2],ymm3[3],ymm1[4,5],ymm3[6],ymm1[7,8,9,10],ymm3[11],ymm1[12,13],ymm3[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm13, %ymm18
-; AVX512F-SLOW-NEXT: vprold $16, %ymm2, %ymm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = 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]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm2
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
-; AVX512F-SLOW-NEXT: vprold $16, %xmm5, %xmm5
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm11[0],xmm7[0],xmm11[1],xmm7[1],xmm11[2],xmm7[2],xmm11[3],xmm7[3]
-; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm2 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm7, %xmm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm11[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm5[0],xmm0[1],xmm5[2,3],xmm0[4],xmm5[5,6],xmm0[7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm22, %ymm11
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm11[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0,1,2],ymm0[3],ymm5[4,5],ymm0[6],ymm5[7,8,9,10],ymm0[11],ymm5[12,13],ymm0[14],ymm5[15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm4, %xmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm6, %xmm3
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm5
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm6, %xmm6
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm6, %ymm23
-; AVX512F-SLOW-NEXT: vprold $16, %xmm5, %xmm6
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2],xmm7[3,4],xmm6[5],xmm7[6,7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm6, %ymm24
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm4
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm4, %xmm2
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm5
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm6[0],xmm2[1],xmm6[2,3],xmm2[4],xmm6[5,6],xmm2[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm28
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm22, %ymm2
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm26
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm22
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqa %ymm9, %ymm7
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm20[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1],ymm15[2],ymm4[3,4],ymm15[5],ymm4[6,7,8,9],ymm15[10],ymm4[11,12],ymm15[13],ymm4[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm17
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm1[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm29[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm6 = mem[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm27[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm19[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm18[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm16[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm13[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm7[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm13[1],ymm1[2,3],ymm13[4],ymm1[5,6,7,8],ymm13[9],ymm1[10,11],ymm13[12],ymm1[13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,0,1,1]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm16 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm3, %zmm20, %zmm16
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm3[4,5,6,7]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm0
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm3, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm30, %zmm11, %zmm16
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm16
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm31, %zmm0, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm11[0,1,2,3],zmm3[4,5,6,7]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm3
-; AVX512F-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm16 = mem[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm15 = xmm12[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm18 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm19 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm21 = mem[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm12 = mem[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,0,1,1]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm28, %xmm11
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm11 = xmm11[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm26[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm22[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm17[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm4, %zmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm8[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm8[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm27, %zmm9
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm9
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm13[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm14[2,1,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm2
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm7, %zmm7
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm15[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm9, %zmm9
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm9
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm10[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm19, %zmm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm21, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm7, %zmm4, %zmm10
-; AVX512F-SLOW-NEXT: vpbroadcastd (%rax), %ymm4
-; AVX512F-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm4
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm4
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm5
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm5
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm24, %zmm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm25, %zmm9
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm7, %zmm20, %zmm9
-; AVX512F-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm10
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm7
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm7
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm30, %zmm29, %zmm9
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm31, %zmm1
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm1
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm6
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm6, 128(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm3, 320(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm7, 256(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm5, 192(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, 64(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, 384(%rax)
-; AVX512F-SLOW-NEXT: addq $824, %rsp # imm = 0x338
-; AVX512F-SLOW-NEXT: vzeroupper
-; AVX512F-SLOW-NEXT: retq
+; AVX512F-ONLY-SLOW-LABEL: store_i16_stride7_vf32:
+; AVX512F-ONLY-SLOW: # %bb.0:
+; AVX512F-ONLY-SLOW-NEXT: subq $648, %rsp # imm = 0x288
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512F-ONLY-SLOW-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>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm4, %ymm30
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm5, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm5, %ymm31
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm15
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm2, %ymm15, %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm6, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm5, %ymm7, %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm5, %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm11, %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm4, %ymm9, %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm2, %ymm4, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm0, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm3, %zmm2, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm2, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm1, %zmm2, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm7 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm9, %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,16,u,u,17,17,u,u,0,u,u,1,2,u,u,3>
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm3, %zmm1, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm12
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm1[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm26 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm8, %zmm1, %zmm26
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm15, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm8[0,1],ymm1[2],ymm8[3,4],ymm1[5],ymm8[6,7,8,9],ymm1[10],ymm8[11,12],ymm1[13],ymm8[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5],ymm6[6],ymm8[7,8,9,10],ymm6[11],ymm8[12,13],ymm6[14],ymm8[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
+; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm1, %zmm6, %zmm28
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm1, %xmm25
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rax), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm6, %ymm8, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm3[4],xmm12[5],xmm3[5],xmm12[6],xmm3[6],xmm12[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rax), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm3, %ymm8, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm6, %ymm1, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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]
+; AVX512F-ONLY-SLOW-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm13, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm6[2],ymm3[3,4],ymm6[5],ymm3[6,7,8,9],ymm6[10],ymm3[11,12],ymm6[13],ymm3[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm13[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm6[1],ymm3[2,3],ymm6[4],ymm3[5,6,7,8],ymm6[9],ymm3[10,11],ymm6[12],ymm3[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm3[0,1,2],ymm5[3],ymm3[4,5],ymm5[6],ymm3[7,8,9,10],ymm5[11],ymm3[12,13],ymm5[14],ymm3[15]
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm4, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm4[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm23
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [6,5,0,0,7,6,0,7,6,5,0,0,7,6,0,7]
+; AVX512F-ONLY-SLOW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermd %zmm1, %zmm0, %zmm29
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu (%rsp), %ymm11 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm11[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm31[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm1, %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm6[0,1],xmm3[2],xmm6[3,4],xmm3[5],xmm6[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm0, %xmm20
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm1, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0],xmm0[1],xmm7[2,3],xmm0[4],xmm7[5,6],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm12, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm3, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm7, %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm6, %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm7 = xmm3[0,1],xmm6[2],xmm3[3,4],xmm6[5],xmm3[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm11[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm30[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm11, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm30[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm15 = ymm3[0,1],ymm12[2],ymm3[3,4],ymm12[5],ymm3[6,7,8,9],ymm12[10],ymm3[11,12],ymm12[13],ymm3[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm14[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm31[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm12[2],ymm3[3,4],ymm12[5],ymm3[6,7,8,9],ymm12[10],ymm3[11,12],ymm12[13],ymm3[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm31[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm14[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3],ymm14[4],ymm12[5,6,7,8],ymm14[9],ymm12[10,11],ymm14[12],ymm12[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm25, %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm1[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm0[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm22[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm17[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm16[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm30
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm30, %zmm13, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm13, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm5 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm8 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm31 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm31, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm2 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm5 = mem[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm27[2,2,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm24[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm23[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm17 = mem[2,3,3,3,6,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm21[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm20, %xmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm11 = xmm9[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm19[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm18, %xmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm8, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm8[0,1,2,3],zmm2[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm14, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm29
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm17[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm22, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm30, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm4, %zmm31, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd (%rax), %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512F-ONLY-SLOW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermd (%rax), %zmm5, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, 320(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, 256(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm29, 384(%rax)
+; AVX512F-ONLY-SLOW-NEXT: addq $648, %rsp # imm = 0x288
+; AVX512F-ONLY-SLOW-NEXT: vzeroupper
+; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride7_vf32:
; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $744, %rsp # imm = 0x2E8
+; AVX512F-ONLY-FAST-NEXT: subq $232, %rsp
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm1, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm1, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm4
; AVX512F-ONLY-FAST-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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm4, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, (%rsp) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm6
; AVX512F-ONLY-FAST-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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm6, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm6, %ymm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm13
; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm14, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm6
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm15
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm6, %ymm7, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm11, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm6, %ymm7, %ymm25
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm6, %ymm5
; AVX512F-ONLY-FAST-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm4, %ymm31
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm1
; AVX512F-ONLY-FAST-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm1, %ymm3, %ymm19
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm10[4],xmm6[5],xmm10[5],xmm6[6],xmm10[6],xmm6[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm10[0],xmm6[1],xmm10[1],xmm6[2],xmm10[2],xmm6[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,1,1,3,8,8,9,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm1, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm4, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm6, %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm6, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [0,0,0,1,8,9,9,11]
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm7, %zmm6, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,0,1,1,12,13,u,15>
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm4, %zmm3, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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]
-; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm1, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm14, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [2,2,2,3,8,8,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm14, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[1,2,2,3,5,6,6,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm14, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[0,0,2,1,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7,8,9,10],ymm5[11],ymm4[12,13],ymm5[14],ymm4[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,2,3,3,10,9,11,10]
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm4, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm5, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm5, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 8(%rax), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpandn %ymm5, %ymm9, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm12[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm13[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[0,0,2,1,4,4,6,5]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm24 = <0,1,u,3,10,10,11,11>
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm19, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,4,u,u,u,5,u>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpandn %ymm3, %ymm4, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm30
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} 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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm31
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm12, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm0, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm2[1,2,2,3,5,6,6,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rax), %zmm29
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
-; AVX512F-ONLY-FAST-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm29, %zmm9, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm3, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm6[0,1],xmm1[2],xmm6[3,4],xmm1[5],xmm6[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm8, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm1, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm4, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm10, %ymm29
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,2,2,3,10,u,11,u>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm22
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq {{.*#+}} xmm9 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm11[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0],xmm12[1],xmm13[2,3],xmm12[4],xmm13[5,6],xmm12[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm6, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0],xmm9[1],xmm12[2,3],xmm9[4],xmm12[5,6],xmm9[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm14, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = <u,2,2,3,8,u,9,u>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm2, %zmm1, %zmm23
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm5, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [2,1,3,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm3, %zmm2, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [2,2,2,3,8,8,8,9]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm7, %xmm24
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm5, %zmm2, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm2, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = <u,0,u,1,8,8,9,u>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm0, %zmm26
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,0,0,1,8,9,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm27
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm13, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd {{.*#+}} ymm5 = [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]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm13, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm15[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm7[3],ymm3[4,5],ymm7[6],ymm3[7,8,9,10],ymm7[11],ymm3[12,13],ymm7[14],ymm3[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm28
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm9
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm0, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm8, %xmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = <0,0,1,1,12,13,u,15>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
+; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 8(%rax), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpandn %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm3, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm3, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm29, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm8 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm31, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm7 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm14, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm14, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm4[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7,8,9,10],ymm5[11],ymm1[12,13],ymm5[14],ymm1[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm29 = <0,1,u,3,10,10,11,11>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm1, %zmm21, %zmm29
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,4,u,u,u,5,u>
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm6, %ymm1, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpandn %ymm1, %ymm5, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm13, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm13 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [2,2,2,3,8,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm21, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm12[2],ymm2[3,4],ymm12[5],ymm2[6,7,8,9],ymm12[10],ymm2[11,12],ymm12[13],ymm2[14,15]
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm3, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm12 = ymm4[0,1],ymm12[2],ymm4[3,4],ymm12[5],ymm4[6,7,8,9],ymm12[10],ymm4[11,12],ymm12[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [2,1,3,2,10,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm31, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm22, %zmm17, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
+; AVX512F-ONLY-FAST-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm3, %zmm6, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm10
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,0,1,1,8,8,10,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm11, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm6 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm6 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm13 = xmm1[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm18, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = [0,1,1,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm22, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm0, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm6[2],xmm2[3,4],xmm6[5],xmm2[6,7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm12[0],xmm6[1],xmm12[2,3],xmm6[4],xmm12[5,6],xmm6[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm2[4],xmm11[5],xmm2[5],xmm11[6],xmm2[6],xmm11[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm22, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 36(%rax), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 40(%rax), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm6, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm6
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm26, %zmm11, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm10, %ymm0
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm5[1],ymm9[2,3],ymm5[4],ymm9[5,6,7,8],ymm5[9],ymm9[10,11],ymm5[12],ymm9[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm5, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm4, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm2, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm11 = ymm4[0,1],ymm2[2],ymm4[3,4],ymm2[5],ymm4[6,7,8,9],ymm2[10],ymm4[11,12],ymm2[13],ymm4[14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm6, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm0[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm2[0,1],ymm15[2],ymm2[3,4],ymm15[5],ymm2[6,7,8,9],ymm15[10],ymm2[11,12],ymm15[13],ymm2[14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm3[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm15 = ymm14[0],ymm15[1],ymm14[2,3],ymm15[4],ymm14[5,6,7,8],ymm15[9],ymm14[10,11],ymm15[12],ymm14[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = <6,u,u,u,7,u,u,7>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm27, %ymm14, %ymm27
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm29, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm16[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm14 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm29 = mem[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm30[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, (%rsp), %ymm3 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm3 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm5 = mem[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm4 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm10[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm10[0,1],ymm5[2],ymm10[3,4],ymm5[5],ymm10[6,7,8,9],ymm5[10],ymm10[11,12],ymm5[13],ymm10[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm12[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm31, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd (%rax), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 4(%rax), %ymm12
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm12[0],ymm2[1],ymm12[2,3],ymm2[4],ymm12[5,6,7,8],ymm2[9],ymm12[10,11],ymm2[12],ymm12[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm1[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm21, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <6,u,u,u,7,u,u,7>
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm16, %ymm2, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512F-ONLY-FAST-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm3, %zmm5, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm5 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm5, %zmm11, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm23, %zmm17, %zmm19
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm30
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm8, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm24, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
-; AVX512F-ONLY-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm1 = mem[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm26[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm17[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm21[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm20[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm22[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm25[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm18[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm30
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm30, %zmm3, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm7
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 36(%rax), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 40(%rax), %ymm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm23, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm29, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd (%rax), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 4(%rax), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 32(%rax), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm27, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512F-ONLY-FAST-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm0, %zmm4, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm2
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm19, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm28, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $744, %rsp # imm = 0x2E8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 384(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm30, 64(%rax)
+; AVX512F-ONLY-FAST-NEXT: addq $232, %rsp
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
+; AVX512DQ-SLOW-LABEL: store_i16_stride7_vf32:
+; AVX512DQ-SLOW: # %bb.0:
+; AVX512DQ-SLOW-NEXT: subq $648, %rsp # imm = 0x288
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm4, %ymm30
+; AVX512DQ-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = [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-SLOW-NEXT: vpshufb %ymm4, %ymm2, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = <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-SLOW-NEXT: vpshufb %ymm9, %ymm5, %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm5, %ymm31
+; AVX512DQ-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm15
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm2, %ymm15, %ymm5
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm6
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = <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-SLOW-NEXT: vpshufb %ymm3, %ymm6, %ymm7
+; AVX512DQ-SLOW-NEXT: vpor %ymm5, %ymm7, %ymm5
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm7
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm8
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm5
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm5, %ymm4
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm9, %ymm11, %ymm9
+; AVX512DQ-SLOW-NEXT: vpor %ymm4, %ymm9, %ymm4
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm13, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %ymm14
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm4
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm2, %ymm4, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm0, %ymm2
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm15[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm3, %zmm2, %zmm9
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm10, %xmm2, %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm1, %zmm2, %zmm3
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm2
+; AVX512DQ-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm7 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm7, %xmm9, %xmm1
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1],xmm3[2,3],xmm1[4],xmm3[5,6],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,16,u,u,17,17,u,u,0,u,u,1,2,u,u,3>
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm3, %zmm1, %zmm8
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm12
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm1[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm26 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm8, %zmm1, %zmm26
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm15, %ymm1
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm6[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm8[0,1],ymm1[2],ymm8[3,4],ymm1[5],ymm8[6,7,8,9],ymm1[10],ymm8[11,12],ymm1[13],ymm8[14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5],ymm6[6],ymm8[7,8,9,10],ymm6[11],ymm8[12,13],ymm6[14],ymm8[15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-SLOW-NEXT: vpermi2q %zmm1, %zmm6, %zmm28
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm1, %xmm25
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rax), %ymm8
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = [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-SLOW-NEXT: vpshufb %ymm6, %ymm8, %ymm2
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm12[4],xmm3[4],xmm12[5],xmm3[5],xmm12[6],xmm3[6],xmm12[7],xmm3[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rax), %ymm1
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm8 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm3, %ymm8, %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm6, %ymm1, %ymm6
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm4[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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-SLOW-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm13, %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm22
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm13[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm6[1],ymm3[2,3],ymm6[4],ymm3[5,6,7,8],ymm6[9],ymm3[10,11],ymm6[12],ymm3[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm3[0,1,2],ymm5[3],ymm3[4,5],ymm5[6],ymm3[7,8,9,10],ymm5[11],ymm3[12,13],ymm5[14],ymm3[15]
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm4, %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = 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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm4[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm0, %ymm23
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 = [6,5,0,0,7,6,0,7,6,5,0,0,7,6,0,7]
+; AVX512DQ-SLOW-NEXT: # zmm0 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermd %zmm1, %zmm0, %zmm29
+; AVX512DQ-SLOW-NEXT: vmovdqu (%rsp), %ymm11 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm11[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm9 = 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-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm14[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm31[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm8 = 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-SLOW-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm1, %xmm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm6[0,1],xmm3[2],xmm6[3,4],xmm3[5],xmm6[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm0, %xmm20
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm1
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm7, %xmm1, %xmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0],xmm0[1],xmm7[2,3],xmm0[4],xmm7[5,6],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm10, %xmm12, %xmm2
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm10, %xmm3, %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm10, %xmm7, %xmm10
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm6, %xmm6
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm7 = xmm3[0,1],xmm6[2],xmm3[3,4],xmm6[5],xmm3[6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm11[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm30[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = 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-SLOW-NEXT: vpshufb %ymm15, %ymm11, %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm30[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm15 = ymm3[0,1],ymm12[2],ymm3[3,4],ymm12[5],ymm3[6,7,8,9],ymm12[10],ymm3[11,12],ymm12[13],ymm3[14,15]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm14[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm31[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm12[2],ymm3[3,4],ymm12[5],ymm3[6,7,8,9],ymm12[10],ymm3[11,12],ymm12[13],ymm3[14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm31[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm14[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0],ymm14[1],ymm12[2,3],ymm14[4],ymm12[5,6,7,8],ymm14[9],ymm12[10,11],ymm14[12],ymm12[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm25, %xmm1
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm1[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm0[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm22[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm17[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm16[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm30
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm13 = [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-SLOW-NEXT: vpternlogq $184, %zmm30, %zmm13, %zmm0
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,1,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm9, %zmm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm13, %zmm4
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm31 = [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-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm31, %zmm8
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm2 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm5 = mem[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm27[2,2,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm24[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm23[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpshufd $254, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm17 = mem[2,3,3,3,6,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm21[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm20, %xmm9
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm11 = xmm9[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm19[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm18, %xmm9
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm1
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm8, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm8[0,1,2,3],zmm2[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm2
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm2
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm14, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm29
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm17[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm22, %zmm4
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm30, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm4, %zmm31, %zmm5
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm4
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm8
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm4
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm10, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm5
+; AVX512DQ-SLOW-NEXT: vpbroadcastd (%rax), %ymm7
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm8
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm3, %zmm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm3
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm5 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512DQ-SLOW-NEXT: # zmm5 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermd (%rax), %zmm5, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm5
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, (%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm2, 320(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, 256(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm29, 384(%rax)
+; AVX512DQ-SLOW-NEXT: addq $648, %rsp # imm = 0x288
+; AVX512DQ-SLOW-NEXT: vzeroupper
+; AVX512DQ-SLOW-NEXT: retq
+;
; AVX512DQ-FAST-LABEL: store_i16_stride7_vf32:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $744, %rsp # imm = 0x2E8
+; AVX512DQ-FAST-NEXT: subq $232, %rsp
; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm1, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm1, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm4
; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm4, %ymm11
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [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-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm8, (%rsp) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm6
; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm6, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm6, %ymm18
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm13
; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm2, %ymm14, %ymm6
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm6
; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm15
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <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-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm7
-; AVX512DQ-FAST-NEXT: vporq %ymm6, %ymm7, %ymm20
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm11
-; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm11, %ymm5
+; AVX512DQ-FAST-NEXT: vporq %ymm6, %ymm7, %ymm25
+; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm6
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm6, %ymm5
; AVX512DQ-FAST-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm12
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm12, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm13
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm13, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm4
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm31
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm12
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm1
; AVX512DQ-FAST-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm1
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm3
-; AVX512DQ-FAST-NEXT: vporq %ymm1, %ymm3, %ymm19
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm10[4],xmm6[5],xmm10[5],xmm6[6],xmm10[6],xmm6[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm3
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm10[0],xmm6[1],xmm10[1],xmm6[2],xmm10[2],xmm6[3],xmm10[3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm1, %xmm4
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm4, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm4
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm6, %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm6, %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm7, %zmm6, %zmm10
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm4
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,0,1,1,12,13,u,15>
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm4, %zmm3, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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-FAST-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm1, %xmm1
-; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm14, %ymm3
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm15[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm3, %zmm4
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vprold $16, %ymm14, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[1,2,2,3,5,6,6,7]
-; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: vpshufb %ymm3, %ymm14, %ymm4
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm15[0,0,2,1,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm5[3],ymm4[4,5],ymm5[6],ymm4[7,8,9,10],ymm5[11],ymm4[12,13],ymm5[14],ymm4[15]
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm4, %zmm7
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm1, %ymm5, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm27
-; AVX512DQ-FAST-NEXT: vpbroadcastd 8(%rax), %ymm5
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpandn %ymm5, %ymm9, %ymm5
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm28
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm4 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm12[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm13[0,1,1,3,4,5,5,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm3
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm2[0,0,2,1,4,4,6,5]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm24 = <0,1,u,3,10,10,11,11>
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm19, %zmm24
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = <u,u,4,u,u,u,5,u>
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpandn %ymm3, %ymm4, %ymm3
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm19
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm8[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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm30
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} 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-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm1, %ymm31
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm12, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vprold $16, %ymm0, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm2[1,2,2,3,5,6,6,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 (%rax), %zmm29
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm7 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
-; AVX512DQ-FAST-NEXT: # zmm7 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm29, %zmm9, %zmm7
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm3
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm3, %xmm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm6[0,1],xmm1[2],xmm6[3,4],xmm1[5],xmm6[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm6
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm8, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm26
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm1, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm1, %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm4
+; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm4, %ymm1
+; AVX512DQ-FAST-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[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-FAST-NEXT: vmovdqa64 %ymm10, %ymm29
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = <u,2,2,3,10,u,11,u>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm2, %zmm22
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm10
; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm11
-; AVX512DQ-FAST-NEXT: vpbroadcastq {{.*#+}} xmm9 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm10, %xmm12
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm13 = xmm11[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0],xmm12[1],xmm13[2,3],xmm12[4],xmm13[5,6],xmm12[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm6, %xmm6
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm6[2],xmm1[3,4],xmm6[5],xmm1[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0],xmm9[1],xmm12[2,3],xmm9[4],xmm12[5,6],xmm9[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm14, %ymm5
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = <u,2,2,3,8,u,9,u>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm2, %zmm1, %zmm23
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = 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-FAST-NEXT: vmovdqa 32(%rdi), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm1
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm3, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm5, %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm19 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm3, %zmm2, %zmm19
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm7, %xmm24
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm5, %zmm2, %zmm20
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm2, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = <u,0,u,1,8,8,9,u>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm0, %zmm26
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm1
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm3
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm27
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm13, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
+; AVX512DQ-FAST-NEXT: vpbroadcastd {{.*#+}} ymm5 = [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-FAST-NEXT: vpshufb %ymm5, %ymm13, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm15[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm7[3],ymm3[4,5],ymm7[6],ymm3[7,8,9,10],ymm7[11],ymm3[12,13],ymm7[14],ymm3[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm3, %zmm28
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm15
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm0
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm9
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm25, %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm8, %xmm18
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = <0,0,1,1,12,13,u,15>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
+; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-FAST-NEXT: vpbroadcastd 8(%rax), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpandn %ymm1, %ymm2, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm3, %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm3, %ymm16
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm1, %zmm30
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm29, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm8 = 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-FAST-NEXT: vmovdqa64 %ymm31, %ymm13
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm7 = 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-FAST-NEXT: vmovdqa %ymm14, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm14, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm4[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm5[3],ymm1[4,5],ymm5[6],ymm1[7,8,9,10],ymm5[11],ymm1[12,13],ymm5[14],ymm1[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm29 = <0,1,u,3,10,10,11,11>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm1, %zmm21, %zmm29
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,4,u,u,u,5,u>
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm6
+; AVX512DQ-FAST-NEXT: vpermd %ymm6, %ymm1, %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpandn %ymm1, %ymm5, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm2
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm14
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[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-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
+; AVX512DQ-FAST-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-FAST-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm13, %ymm13
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm13 = 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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm21, %zmm13
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm12[2],ymm2[3,4],ymm12[5],ymm2[6,7,8,9],ymm12[10],ymm2[11,12],ymm12[13],ymm2[14,15]
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm3, %ymm12
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm12 = ymm4[0,1],ymm12[2],ymm4[3,4],ymm12[5],ymm4[6,7,8,9],ymm12[10],ymm4[11,12],ymm12[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm31, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [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-FAST-NEXT: vpternlogq $226, %zmm22, %zmm17, %zmm13
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 (%rax), %zmm3
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm4 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
+; AVX512DQ-FAST-NEXT: # zmm4 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm3, %zmm6, %zmm4
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm4
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm10, %xmm10
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm6[0,2,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm11, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm6 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm6 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm13 = xmm1[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm18, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm22, %zmm13
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm0, %xmm6
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm15[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm6[2],xmm2[3,4],xmm6[5],xmm2[6,7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm12 = xmm11[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm12[0],xmm6[1],xmm12[2,3],xmm6[4],xmm12[5,6],xmm6[7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm2[4],xmm11[5],xmm2[5],xmm11[6],xmm2[6],xmm11[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm22, %zmm2
+; AVX512DQ-FAST-NEXT: vpbroadcastd 36(%rax), %ymm6
+; AVX512DQ-FAST-NEXT: vpbroadcastd 40(%rax), %ymm11
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm6, %zmm6
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm6
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $226, %zmm26, %zmm11, %zmm10
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm6
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm10, %ymm0
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm5[1],ymm9[2,3],ymm5[4],ymm9[5,6,7,8],ymm5[9],ymm9[10,11],ymm5[12],ymm9[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm5, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm0[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5],ymm4[6],ymm5[7,8,9,10],ymm4[11],ymm5[12,13],ymm4[14],ymm5[15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm18
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm9
-; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm2, %xmm12
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm11 = 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-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[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-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm10 = 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-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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-FAST-NEXT: vmovdqa %ymm6, %ymm3
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm15 = ymm0[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm2[0,1],ymm15[2],ymm2[3,4],ymm15[5],ymm2[6,7,8,9],ymm15[10],ymm2[11,12],ymm15[13],ymm2[14,15]
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm3[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm14 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm15 = ymm14[0],ymm15[1],ymm14[2,3],ymm15[4],ymm14[5,6,7,8],ymm15[9],ymm14[10,11],ymm15[12],ymm14[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = <6,u,u,u,7,u,u,7>
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm27, %ymm14, %ymm27
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm29, %zmm0
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm1 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm16[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm14 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm29 = mem[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm30[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm31[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vpermq $234, (%rsp), %ymm3 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm3 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm5 = mem[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm4 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = [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-FAST-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm8
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm28
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm28
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-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-FAST-NEXT: vpunpcklwd {{.*#+}} ymm5 = ymm10[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm5 = ymm10[0,1],ymm5[2],ymm10[3,4],ymm5[5],ymm10[6,7,8,9],ymm5[10],ymm10[11,12],ymm5[13],ymm10[14,15]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm10 = xmm12[0,2,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm31, %zmm5
+; AVX512DQ-FAST-NEXT: vpbroadcastd (%rax), %ymm0
+; AVX512DQ-FAST-NEXT: vpbroadcastd 4(%rax), %ymm12
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm27, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqu (%rsp), %ymm13 # 32-byte Reload
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm12[0],ymm2[1],ymm12[2,3],ymm2[4],ymm12[5,6,7,8],ymm2[9],ymm12[10,11],ymm2[12],ymm12[13,14,15]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm12 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm1[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-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-FAST-NEXT: vpermt2q %zmm2, %zmm21, %zmm12
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = <6,u,u,u,7,u,u,7>
+; AVX512DQ-FAST-NEXT: vpermd %ymm16, %ymm2, %ymm2
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm3
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm5 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512DQ-FAST-NEXT: # zmm5 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermd %zmm3, %zmm5, %zmm3
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm28, %zmm3
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm3
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm9 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm5, %zmm11, %zmm9
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm23, %zmm17, %zmm19
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm25, %zmm30
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm30
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm8, %zmm8
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm8
-; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm24, %zmm19
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm19
-; AVX512DQ-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm1 = mem[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm26[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm17[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm13 = xmm13[0,2,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm21[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm20[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm22[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm24 = ymm25[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm18[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm30, %zmm30
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm30, %zmm3, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm7
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm1, %zmm16, %zmm2
-; AVX512DQ-FAST-NEXT: vpbroadcastd 36(%rax), %ymm1
-; AVX512DQ-FAST-NEXT: vpbroadcastd 40(%rax), %ymm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm9, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm21, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm23, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm29, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm5
-; AVX512DQ-FAST-NEXT: vpbroadcastd (%rax), %ymm2
-; AVX512DQ-FAST-NEXT: vpbroadcastd 4(%rax), %ymm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
-; AVX512DQ-FAST-NEXT: vpbroadcastd 32(%rax), %ymm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm27, %zmm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm3
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm3
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm11, %zmm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm4 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512DQ-FAST-NEXT: # zmm4 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %zmm0, %zmm4, %zmm4
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm4
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm7
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm29, %zmm14
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm14
+; AVX512DQ-FAST-NEXT: vpbroadcastd 32(%rax), %ymm5
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm20, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm2
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm19, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm28, 64(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-FAST-NEXT: addq $744, %rsp # imm = 0x2E8
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, 128(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 256(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 384(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm30, 64(%rax)
+; AVX512DQ-FAST-NEXT: addq $232, %rsp
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
@@ -11085,2169 +11572,2649 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
-; AVX512F-SLOW-LABEL: store_i16_stride7_vf64:
-; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: subq $2456, %rsp # imm = 0x998
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rcx), %ymm12
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdx), %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdi), %ymm8
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rsi), %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm12, %ymm2
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm3
-; AVX512F-SLOW-NEXT: vporq %ymm2, %ymm3, %ymm19
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm4, %ymm2
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm8, %ymm3
-; AVX512F-SLOW-NEXT: vporq %ymm2, %ymm3, %ymm20
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512F-SLOW-NEXT: vmovdqa 64(%r9), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa %ymm5, %ymm14
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512F-SLOW-NEXT: vmovdqa 64(%r8), %ymm5
-; AVX512F-SLOW-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>
-; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm5, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm5, %ymm26
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rcx), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm17
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rdx), %ymm10
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm10, %ymm3
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %ymm9
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm9, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rdi), %ymm6
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm6, %ymm3
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm3
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm30
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm5, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm5, %ymm28
-; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm5
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm5, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm3
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm2, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm1
-; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm1, %ymm11
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm11, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm11
-; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm11, %ymm13
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %ymm0
-; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm0, %ymm15
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm14, %ymm16
-; AVX512F-SLOW-NEXT: vpor %ymm15, %ymm13, %ymm13
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vprold $16, %ymm0, %ymm13
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm15[0,1],ymm13[2],ymm15[3,4],ymm13[5],ymm15[6,7,8,9],ymm13[10],ymm15[11,12],ymm13[13],ymm15[14,15]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4],ymm14[5],ymm15[6,7,8,9],ymm14[10],ymm15[11,12],ymm14[13],ymm15[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [2,1,3,2,10,10,10,11]
-; AVX512F-SLOW-NEXT: vpermi2q %zmm14, %zmm13, %zmm15
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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]
-; AVX512F-SLOW-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm12, %ymm13
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm14, %ymm27
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3,4],ymm14[5],ymm13[6,7,8,9],ymm14[10],ymm13[11,12],ymm14[13],ymm13[14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3],ymm13[4],ymm14[5,6,7,8],ymm13[9],ymm14[10,11],ymm13[12],ymm14[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,2,2,3]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm4[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm8[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3],ymm13[4],ymm14[5,6,7,8],ymm13[9],ymm14[10,11],ymm13[12],ymm14[13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,2,2,3]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm13 = ymm4[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm8[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0,1,2],ymm13[3],ymm14[4,5],ymm13[6],ymm14[7,8,9,10],ymm13[11],ymm14[12,13],ymm13[14],ymm14[15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,1,3,3]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 96(%r9), %ymm13
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm13[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[3,3,3,3]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm13[2,1,2,3,6,5,6,7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm14[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,2]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm12[2],ymm7[3,4],ymm12[5],ymm7[6,7,8,9],ymm12[10],ymm7[11,12],ymm12[13],ymm7[14,15]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm4 = ymm4[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
-; AVX512F-SLOW-NEXT: vprold $16, %ymm13, %ymm8
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm16, %ymm12
-; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm13, %ymm12
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm19, %zmm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm20, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm4
-; AVX512F-SLOW-NEXT: vmovdqa 96(%r8), %ymm7
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm7[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm7[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[14,15],zero,zero,ymm7[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm7[16,17],zero,zero,ymm7[u,u],zero,zero
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm14, %ymm4, %ymm7
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm12
-; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm4, %ymm4
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm13[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm7
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm8[2,2,2,2]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm4
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm4
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm12[0,1,2,3],zmm4[4,5,6,7]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rax), %ymm4
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm8 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm4, %ymm4
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpandn %ymm7, %ymm14, %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm4
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpbroadcastd 72(%rax), %ymm4
-; AVX512F-SLOW-NEXT: vpandnq %ymm4, %ymm25, %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rax), %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm4, %ymm12
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm7
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 64(%r9), %xmm13
-; AVX512F-SLOW-NEXT: vmovdqa 64(%r8), %xmm12
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm13, %xmm16
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-SLOW-NEXT: vpshufb %xmm13, %xmm7, %xmm7
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm13, %xmm19
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vpandnq %ymm7, %ymm25, %ymm13
-; AVX512F-SLOW-NEXT: vmovdqa (%rax), %ymm7
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm7, %ymm15
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm13, %zmm13
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm3[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm15[0,1],ymm13[2],ymm15[3,4],ymm13[5],ymm15[6,7,8,9],ymm13[10],ymm15[11,12],ymm13[13],ymm15[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm2[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm1[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4],ymm15[5],ymm13[6,7,8,9],ymm15[10],ymm13[11,12],ymm15[13],ymm13[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm13, %ymm22
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rax), %ymm15
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm15[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpandn %ymm13, %ymm14, %ymm13
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm15, %ymm8
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm15, %ymm29
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm8
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3],ymm0[4,5],ymm8[6],ymm0[7,8,9,10],ymm8[11],ymm0[12,13],ymm8[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm27, %ymm11
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm5, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm8[2],ymm0[3,4],ymm8[5],ymm0[6,7,8,9],ymm8[10],ymm0[11,12],ymm8[13],ymm0[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3],ymm3[4],ymm0[5,6,7,8],ymm3[9],ymm0[10,11],ymm3[12],ymm0[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm2[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm31
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm17, %ymm13
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm13[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm13, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm18, %ymm5
-; AVX512F-SLOW-NEXT: vprold $16, %ymm18, %ymm0
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm26[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm26[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1,2],ymm2[3],ymm0[4,5],ymm2[6],ymm0[7,8,9,10],ymm2[11],ymm0[12,13],ymm2[14],ymm0[15]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,2,3,3,10,9,11,10]
-; AVX512F-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,3,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm27 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm27, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm13[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm26[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm5[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
-; AVX512F-SLOW-NEXT: vmovdqa 96(%r9), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa 96(%r8), %xmm2
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermt2d %zmm8, %zmm5, %zmm6
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpbroadcastd 96(%rax), %ymm8
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm6, %zmm18, %zmm4
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rsi), %xmm4
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdi), %xmm6
-; AVX512F-SLOW-NEXT: vprold $16, %xmm4, %xmm8
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2],xmm9[3,4],xmm8[5],xmm9[6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rcx), %xmm4
-; AVX512F-SLOW-NEXT: vmovdqa 96(%rdx), %xmm6
-; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm15 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm4, %xmm8
-; AVX512F-SLOW-NEXT: vmovdqa %xmm15, %xmm10
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm6[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0],xmm8[1],xmm9[2,3],xmm8[4],xmm9[5,6],xmm8[7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm8, (%rsp) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm19, %xmm26
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm19, %xmm3
-; AVX512F-SLOW-NEXT: vpshufb %xmm3, %xmm1, %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
-; AVX512F-SLOW-NEXT: vpermt2d %zmm2, %zmm3, %zmm1
-; AVX512F-SLOW-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512F-SLOW-NEXT: vpbroadcastd 104(%rax), %ymm3
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rdi), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm15, %xmm17
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512F-SLOW-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rcx), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm1, %xmm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3],xmm1[4],xmm2[5,6],xmm1[7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
-; AVX512F-SLOW-NEXT: vpermt2d %zmm2, %zmm3, %zmm1
-; AVX512F-SLOW-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512F-SLOW-NEXT: vpbroadcastd 68(%rax), %ymm6
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
-; AVX512F-SLOW-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3,4],xmm2[5],xmm1[6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm6
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm1, %xmm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm6[0],xmm1[1],xmm6[2,3],xmm1[4],xmm6[5,6],xmm1[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm25
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm10
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm10 = xmm1[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermt2d %zmm10, %zmm3, %zmm1
-; AVX512F-SLOW-NEXT: vpbroadcastd (%rax), %ymm3
-; AVX512F-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm10
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm3, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm30, %ymm12
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm12[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm3[2],ymm1[3,4],ymm3[5],ymm1[6,7,8,9],ymm3[10],ymm1[11,12],ymm3[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm12[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm28, %ymm16
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3],ymm1[4],ymm3[5,6,7,8],ymm1[9],ymm3[10,11],ymm1[12],ymm3[13,14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm20
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vprold $16, %ymm2, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm10[1,2,2,3,5,6,6,7]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm10[0,0,2,1,4,4,6,5]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm9[3],ymm3[4,5],ymm9[6],ymm3[7,8,9,10],ymm9[11],ymm3[12,13],ymm9[14],ymm3[15]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpermt2q %zmm1, %zmm0, %zmm3
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm7[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm28
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm27, %zmm28
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = 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]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm3[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm12[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm2[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm3[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermt2d %zmm9, %zmm5, %zmm1
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm30
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm18, %zmm30
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512F-SLOW-NEXT: vprold $16, %xmm7, %xmm9
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm1[1,1,2,3]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm12[0,1],xmm9[2],xmm12[3,4],xmm9[5],xmm12[6,7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3]
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm26, %xmm2
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm6, %xmm6
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm11, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm15, %xmm1
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm14, %xmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm9, %xmm5
-; AVX512F-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm4, %xmm12
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm17, %xmm15
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm4 = xmm15[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm4[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm6[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm6[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm13[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm6[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm27 = mem[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[2,2,2,3]
-; AVX512F-SLOW-NEXT: vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm31[2,1,3,3]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,6]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
-; AVX512F-SLOW-NEXT: vpermt2d %zmm3, %zmm2, %zmm12
-; AVX512F-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm3
-; AVX512F-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm7
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm31
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm31
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm12 = xmm3[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm14
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm15[0],xmm12[1],xmm15[2,3],xmm12[4],xmm15[5,6],xmm12[7]
-; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm14[0],xmm3[0],xmm14[1],xmm3[1],xmm14[2],xmm3[2],xmm14[3],xmm3[3]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm14[4],xmm3[4],xmm14[5],xmm3[5],xmm14[6],xmm3[6],xmm14[7],xmm3[7]
-; AVX512F-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm18 = ymm29[0,1,2,2,4,5,6,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm17 = ymm29[2,3,3,3,6,7,7,7]
-; AVX512F-SLOW-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[2,2,2,3]
-; AVX512F-SLOW-NEXT: vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm7 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpermpd $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[2,1,3,2]
-; AVX512F-SLOW-NEXT: vmovups %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm29 = mem[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm1[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm13 = mem[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm8 = mem[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm9 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm0[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm11 = mem[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm12[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm25[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm21[2,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm19[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm23[2,2,2,3]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512F-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm1 = mem[2,1,3,2]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm1 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm23 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm4[0,1,2,3],zmm23[4,5,6,7]
-; AVX512F-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm23 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm4 # 32-byte Folded Reload
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm25, %zmm4
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm2, %zmm4
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm26, %zmm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm10 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm25, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm23, %zmm2, %zmm10
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm27, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm2, %zmm26
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm26
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm22 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm0, %zmm2
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm27 = zmm16[0,1,2,3],zmm2[4,5,6,7]
-; AVX512F-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm27 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm1, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm1 = mem[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[2,1,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm16
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm16
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm18[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm17[2,1,3,2]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm17
-; AVX512F-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm17 # 64-byte Folded Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm8[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm2, %zmm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm15, %zmm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm2, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm23 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm23
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm24 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm24
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm20, %zmm21, %zmm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm19, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm19
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm28
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm20 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm20, %zmm1
-; AVX512F-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm0 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshuflw $180, (%rsp), %xmm2 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm2 = mem[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm2 = mem[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm3 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm3 = mem[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm6 = mem[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm8 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm9 = mem[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512F-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm13 = mem[2,1,3,3]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm14 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm15 = mem[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm5 = mem[2,1,2,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,5,4]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,1,3]
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm18 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
-; AVX512F-SLOW-NEXT: # xmm7 = mem[0,2,3,3,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm20, %zmm2
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm12[0,1,1,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm11
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm20, %zmm11
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm12
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm11, %zmm0, %zmm30
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm2
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm25, %zmm2
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm15, %zmm0
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm18, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm25, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm5
-; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm31
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm26, %zmm27
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm17
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm27, 320(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm31, 256(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm30, 192(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm28, 128(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm24, (%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm23, 448(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm5, 704(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm12, 640(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm19, 576(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, 512(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm17, 384(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm22, 768(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm16, 832(%rax)
-; AVX512F-SLOW-NEXT: addq $2456, %rsp # imm = 0x998
-; AVX512F-SLOW-NEXT: vzeroupper
-; AVX512F-SLOW-NEXT: retq
+; AVX512F-ONLY-SLOW-LABEL: store_i16_stride7_vf64:
+; AVX512F-ONLY-SLOW: # %bb.0:
+; AVX512F-ONLY-SLOW-NEXT: subq $2168, %rsp # imm = 0x878
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512F-ONLY-SLOW-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>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm6, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vporq %ymm1, %ymm2, %ymm18
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm9, %ymm1
+; AVX512F-ONLY-SLOW-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>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm7, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vporq %ymm1, %ymm2, %ymm17
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm3, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %ymm3
+; AVX512F-ONLY-SLOW-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>
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %ymm10
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm4, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm8, %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm2, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm2, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm8, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm8, %ymm19
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm2, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm11, %ymm13
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm13
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm13, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm12, %ymm0, %ymm15
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm12, %ymm20
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm15, %ymm14, %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm0, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm13[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4],ymm14[5],ymm15[6,7,8,9],ymm14[10],ymm15[11,12],ymm14[13],ymm15[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm13[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm15[0,1],ymm12[2],ymm15[3,4],ymm12[5],ymm15[6,7,8,9],ymm12[10],ymm15[11,12],ymm12[13],ymm15[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [2,1,3,2,10,10,10,11]
+; AVX512F-ONLY-SLOW-NEXT: vpermi2q %zmm12, %zmm14, %zmm15
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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]
+; AVX512F-ONLY-SLOW-NEXT: # ymm14 = mem[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm16, %ymm15
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm15, %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm14, %ymm16
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7,8,9],ymm14[10],ymm12[11,12],ymm14[13],ymm12[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0],ymm12[1],ymm14[2,3],ymm12[4],ymm14[5,6,7,8],ymm12[9],ymm14[10,11],ymm12[12],ymm14[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0],ymm12[1],ymm14[2,3],ymm12[4],ymm14[5,6,7,8],ymm12[9],ymm14[10,11],ymm12[12],ymm14[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm12, %ymm25
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0,1,2],ymm12[3],ymm14[4,5],ymm12[6],ymm14[7,8,9,10],ymm12[11],ymm14[12,13],ymm12[14],ymm14[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm12[2],ymm6[3,4],ymm12[5],ymm6[6,7,8,9],ymm12[10],ymm6[11,12],ymm12[13],ymm6[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm9[0,1],ymm7[2],ymm9[3,4],ymm7[5],ymm9[6,7,8,9],ymm7[10],ymm9[11,12],ymm7[13],ymm9[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm18, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm17, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = ymm12[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm12[14,15],zero,zero,ymm12[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm12[16,17],zero,zero,ymm12[u,u],zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, %ymm9, %ymm7, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %ymm15
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm20, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm15, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm7, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm15, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm14[0,1,2,3],zmm6[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = [6,5,0,0,7,6,0,7,6,5,0,0,7,6,0,7]
+; AVX512F-ONLY-SLOW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpermd %zmm6, %zmm18, %zmm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm6, %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm14, %ymm9, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm6, %zmm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 72(%rax), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpandnq %ymm6, %ymm17, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rax), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm6, %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm14, %zmm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vpandnq %ymm12, %ymm17, %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rax), %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm7, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm7, %ymm22
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm12
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm1[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7,8,9],ymm14[10],ymm12[11,12],ymm14[13],ymm12[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rax), %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm14, %ymm9, %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm12, %ymm14
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm14, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm13[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm9[3],ymm0[4,5],ymm9[6],ymm0[7,8,9,10],ymm9[11],ymm0[12,13],ymm9[14],ymm0[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm16, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm8, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7,8,9],ymm9[10],ymm0[11,12],ymm9[13],ymm0[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = 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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm31
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpermd %zmm12, %zmm18, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm7, %ymm10, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm21, %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm21, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm9[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm23[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,2,3,3,10,9,11,10]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm0, %zmm7, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512F-ONLY-SLOW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermd 64(%rax), %zmm15, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm14, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm23[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm9[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm1, %zmm11, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[2,3,3,3,6,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 96(%rax), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm1, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rsi), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdi), %xmm5
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm4, %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm6, %xmm23
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rcx), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 96(%rdx), %xmm5
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm13 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm4, %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0],xmm6[1],xmm9[2,3],xmm6[4],xmm9[5,6],xmm6[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm30 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm3, %zmm30, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 100(%rax), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 104(%rax), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm28, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rcx), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdx), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm5 = <16,16,17,17,17,17,u,u,0,1,0,1,2,3,2,3>
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm3, %zmm5, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rdi), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%rsi), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm6, %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm9, %xmm27
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm4, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4],xmm4[5],xmm3[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm4
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 64(%r8), %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm4, %zmm6, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 64(%rax), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 68(%rax), %ymm9
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm4, %zmm4
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm9, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm4, %xmm26
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm10[0],xmm2[1],xmm10[2,3],xmm2[4],xmm10[5,6],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm10 = xmm12[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm10, %zmm5, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm12, %xmm12
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm4, %xmm21
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm10, %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm10[2],xmm5[3,4],xmm10[5],xmm5[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm12, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm3, %zmm6, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd (%rax), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm9, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm24, %ymm10
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm19[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %ymm3, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm9[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm9[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm6[3],ymm5[4,5],ymm6[6],ymm5[7,8,9,10],ymm6[11],ymm5[12,13],ymm6[14],ymm5[15]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2q %zmm2, %zmm7, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm10[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm19[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3],ymm2[4],ymm6[5,6,7,8],ymm2[9],ymm6[10,11],ymm2[12],ymm6[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512F-ONLY-SLOW-NEXT: vpermd (%rax), %zmm15, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm14, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3],ymm5[4],ymm2[5,6,7,8],ymm5[9],ymm2[10,11],ymm5[12],ymm2[13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm9[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm3[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,3,6,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm12, %zmm11, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm22[2,3,3,3,6,7,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm12
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm22
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm22
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vprold $16, %xmm2, %xmm12
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm1[1,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm15[0,1],xmm12[2],xmm15[3,4],xmm12[5],xmm15[6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm8, %xmm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm23, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm24 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm19 = mem[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm8 = mem[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,5,7,6]
+; AVX512F-ONLY-SLOW-NEXT: vpermt2d %zmm1, %zmm30, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm30
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm28, %zmm30
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm0, %xmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[1,1,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0],xmm5[1],xmm15[2,3],xmm5[4],xmm15[5,6],xmm5[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm0[3,3,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm23 = mem[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm13[2,2,2,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm2[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm2[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm27, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm3[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm26, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm3[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm5[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm5 = mem[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm10 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm11 = mem[2,2,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm0 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm29 = mem[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm26 = mem[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm21 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm15 = mem[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm4 = mem[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm12 = mem[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm24, %zmm24
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm25, %zmm0
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm24, %zmm27, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm23[2,1,3,2]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm28
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm8 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm23 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm23[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm23 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm13 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm13, %zmm24, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm8, %zmm9, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm8 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm24, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm23, %zmm9, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm18
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm18
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm19 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm23 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm2, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm28[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm25
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm28 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm4, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm17, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm20
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm21
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm31, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm27, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm0 = mem[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm3 = mem[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm3 = mem[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm5 = mem[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm5 = mem[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm6 = mem[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm26 = mem[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm11 = mem[0,2,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm14 = mem[2,1,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm15 = mem[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm16 = mem[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm2 = mem[2,1,2,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # ymm17 = mem[0,0,1,1]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
+; AVX512F-ONLY-SLOW-NEXT: # xmm4 = mem[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm27, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm12[0,1,1,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm11
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm27, %zmm11
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm12
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm11, %zmm0, %zmm22
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm26, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm24, %zmm3
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm16, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm17, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm24, %zmm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm4
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm30
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm23
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm28, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm23, 320(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm30, 256(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, 192(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm21, 128(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, (%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm13, 448(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, 704(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm12, 640(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, 576(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 512(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, 384(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm19, 768(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm25, 832(%rax)
+; AVX512F-ONLY-SLOW-NEXT: addq $2168, %rsp # imm = 0x878
+; AVX512F-ONLY-SLOW-NEXT: vzeroupper
+; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: store_i16_stride7_vf64:
; AVX512F-ONLY-FAST: # %bb.0:
-; AVX512F-ONLY-FAST-NEXT: subq $2264, %rsp # imm = 0x8D8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm2
-; AVX512F-ONLY-FAST-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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm2, %ymm3, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm4, %ymm2
-; AVX512F-ONLY-FAST-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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm10, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm2, %ymm3, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: subq $1432, %rsp # imm = 0x598
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm4, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %ymm5
+; AVX512F-ONLY-FAST-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>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm6, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm7, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm1, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm2, %ymm3, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm6, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm2, %ymm3, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm15, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm5, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm1
; AVX512F-ONLY-FAST-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpor %ymm0, %ymm11, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm14, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vporq %ymm12, %ymm11, %ymm22
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm11 = 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]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm1, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [2,1,3,2,10,10,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm11, %zmm12, %zmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm4, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm12, %ymm18
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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]
-; AVX512F-ONLY-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} 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>
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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]
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm30
+; AVX512F-ONLY-FAST-NEXT: vpor %ymm13, %ymm9, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm9, %ymm13, %ymm17
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpor %ymm9, %ymm13, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm13
+; AVX512F-ONLY-FAST-NEXT: vpor %ymm9, %ymm13, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpor %ymm9, %ymm11, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %ymm11
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm11, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %ymm15
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm15, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vporq %ymm9, %ymm8, %ymm18
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[0,1,1,3,4,5,5,7]
; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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]
-; AVX512F-ONLY-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm0, %ymm31
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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]
+; AVX512F-ONLY-FAST-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm12, %ymm31
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-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]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm16, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm17, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm19, %zmm9
; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <5,u,u,u,6,u,u,6>
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,4,u,u,u,5,u,u>
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15],zero,zero,ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17],zero,zero,ymm10[u,u],zero,zero
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %ymm23, %ymm9, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm0, %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm8[14,15],zero,zero,ymm8[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm8[16,17],zero,zero,ymm8[u,u],zero,zero
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %ymm13, %ymm9, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm13, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %ymm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm13, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm12, %ymm10
; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm0, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,4,u,u,u,5,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm12, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm13, %ymm9
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm11[0,1,2,3],zmm8[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm12, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 64(%rax), %zmm26
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
-; AVX512F-ONLY-FAST-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rax), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm26, %zmm8, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rax), %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm9, %ymm10, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vpandn %ymm10, %ymm14, %ymm10
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm9, %ymm12
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm12, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm12, %ymm23
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5,6,7,8],ymm6[9],ymm7[10,11],ymm6[12],ymm7[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = [0,2,2,3,10,9,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm25, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u>
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm6, %ymm26
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm6, %ymm29
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm20, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm7, %zmm5, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <5,u,u,u,6,u,u,6>
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm4, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 64(%rax), %zmm28
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
+; AVX512F-ONLY-FAST-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm28, %zmm9, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 72(%rax), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpandn %ymm4, %ymm7, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm7, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rax), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm13, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm5, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm5, %ymm27
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm12
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm5, %xmm21
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm7, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm13, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [0,0,1,1,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %zmm4, %zmm7, %zmm9
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [0,1,4,5,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm8, %ymm24, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vpandnq %ymm9, %ymm23, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm8, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 72(%rax), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vpandn %ymm8, %ymm9, %ymm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm9, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rax), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm11, %ymm9
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm16
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r9), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%r8), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,0,1,1,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm10, %zmm21, %zmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %zmm27, %zmm13, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 8(%rax), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpandnq %ymm7, %ymm16, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm6, (%rsp) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm14, %xmm14
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm13, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %zmm4, %zmm14, %zmm16
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 8(%rax), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpandnq %ymm10, %ymm17, %ymm17
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rax), %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm10, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm10, %ymm29
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm17, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r9), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%r8), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm7[4],xmm4[4],xmm7[5],xmm4[5],xmm7[6],xmm4[6],xmm7[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm20, %zmm21, %zmm13
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, %zmm27, %zmm13, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm5[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm31, %ymm4
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm3[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd {{.*#+}} ymm9 = [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]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm7[3],ymm4[4,5],ymm7[6],ymm4[7,8,9,10],ymm7[11],ymm4[12,13],ymm7[14],ymm4[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = <0,1,u,3,10,10,11,11>
+; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm4, %zmm18, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-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]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm10[3],ymm0[4,5],ymm10[6],ymm0[7,8,9,10],ymm10[11],ymm0[12,13],ymm10[14],ymm0[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = <0,1,u,3,10,10,11,11>
-; AVX512F-ONLY-FAST-NEXT: vpermi2q %zmm0, %zmm22, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpandnq %ymm10, %ymm23, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm12
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm28, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm2, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = 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]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm18, %ymm14
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm25, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm14
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm4
; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm15, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rax), %zmm16
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [14,21,0,0,15,22,0,15,14,21,0,0,15,22,0,15]
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm0, %zmm16, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm13, %ymm6, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm7[0,1,2],ymm4[3],ymm7[4,5],ymm4[6],ymm7[7,8,9,10],ymm4[11],ymm7[12,13],ymm4[14],ymm7[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm23, %ymm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm13[0],ymm7[1],ymm13[2,3],ymm7[4],ymm13[5,6,7,8],ymm7[9],ymm13[10,11],ymm7[12],ymm13[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm25, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm26, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm3, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm29, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm20, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm7, %zmm10, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm4 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm15, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [2,1,3,2,10,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm30[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm31, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm6
; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm2, %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %ymm4, %ymm13
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm3, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[1,2,2,3,5,6,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm7, %ymm3, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [2,2,3,3,10,9,11,10]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm12, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm26, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512F-ONLY-FAST-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm0, %zmm1, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm6, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [2,2,2,3,8,8,8,9]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [6,7,3,3,7,7,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm11, %ymm3, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 96(%rax), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm3
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm2, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm3[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2],xmm7[3,4],xmm6[5],xmm7[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm6, %xmm24
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastq {{.*#+}} xmm13 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm2, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0],xmm3[1],xmm7[2,3],xmm3[4],xmm7[5,6],xmm3[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm17, %xmm31
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm17, %xmm4
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm4, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,1,1,3,8,8,9,9]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 104(%rax), %ymm4
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm30[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm7[2],ymm6[3,4],ymm7[5],ymm6[6,7,8,9],ymm7[10],ymm6[11,12],ymm7[13],ymm6[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm20, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm1, %ymm0, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm29, %ymm31
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7,8,9],ymm7[10],ymm4[11,12],ymm7[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm11[0,1],ymm7[2],ymm11[3,4],ymm7[5],ymm11[6,7,8,9],ymm7[10],ymm11[11,12],ymm7[13],ymm11[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm17, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm6, %zmm23, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm15, %ymm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm19[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm19[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm11[3],ymm6[4,5],ymm11[6],ymm6[7,8,9,10],ymm11[11],ymm6[12,13],ymm11[14],ymm6[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = [2,2,3,3,10,9,11,10]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm25, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm28, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512F-ONLY-FAST-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm4, %zmm29, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm6, %zmm28, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm15, %xmm30
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm13, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2,3],xmm0[4],xmm2[5,6],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rcx), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdx), %xmm11
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,2,2,3,8,9,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm0, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm5 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdx), %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0],xmm0[1],xmm3[2,3],xmm0[4],xmm3[5,6],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm3, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm3, %zmm13
; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rdi), %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm13, %xmm19
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2],xmm0[3,4],xmm2[5],xmm0[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm5, %xmm9
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm5, %xmm0, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm14, %xmm11
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,0,1,8,9,9,11]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm5, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 68(%rax), %ymm6
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm2, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm10, %xmm6, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3],xmm6[4],xmm7[5,6],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 64(%rsi), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm3, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2],xmm6[3,4],xmm4[5],xmm6[6,7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [0,0,1,1,8,8,10,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm26, %zmm14
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm3, %xmm12
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,0,0,1,8,9,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm21, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm3[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3],xmm7[4],xmm8[5,6],xmm7[7]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm2, %xmm24
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm13, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsi), %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm8, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm8[2],xmm6[3,4],xmm8[5],xmm6[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm6, %ymm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # xmm6 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm9, %xmm22
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm11, %xmm18
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm8, %zmm5, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 4(%rax), %ymm8
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm6, %zmm2, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm6[2],ymm2[3,4],ymm6[5],ymm2[6,7,8,9],ymm6[10],ymm2[11,12],ymm6[13],ymm2[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm26
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3],ymm2[4],ymm6[5,6,7,8],ymm2[9],ymm6[10,11],ymm2[12],ymm6[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm5, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm8, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm2[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm3[2],xmm7[3,4],xmm3[5],xmm7[6,7]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm9, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm26, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 64(%rax), %ymm1
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 68(%rax), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm6, %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm21, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 4(%rax), %ymm6
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rsi), %xmm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%rdi), %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [2,1,3,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm14, %zmm2, %zmm18
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm3, %zmm2, %zmm13
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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]
+; AVX512F-ONLY-FAST-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm3, %ymm4, %ymm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm4, %ymm16
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm20[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [2,2,2,3,8,10,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm31, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb %ymm2, %ymm15, %ymm2
; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[1,2,2,3,5,6,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0,1],ymm2[2],ymm6[3,4],ymm2[5],ymm6[6,7,8,9],ymm2[10],ymm6[11,12],ymm2[13],ymm6[14,15]
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm7[0,0,2,1,4,4,6,5]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1,2],ymm8[3],ymm6[4,5],ymm8[6],ymm6[7,8,9,10],ymm8[11],ymm6[12,13],ymm8[14],ymm6[15]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm12, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm9[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0,1],ymm2[2],ymm6[3,4],ymm2[5],ymm6[6,7,8,9],ymm2[10],ymm6[11,12],ymm2[13],ymm6[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm6[2],ymm2[3,4],ymm6[5],ymm2[6,7,8,9],ymm6[10],ymm2[11,12],ymm6[13],ymm2[14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm2, %ymm21
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm16, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512F-ONLY-FAST-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermd %zmm2, %zmm6, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm8, %zmm1, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm9[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm9 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm7, %ymm21
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm14 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7,8,9],ymm3[10],ymm4[11,12],ymm3[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm17, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [2,2,2,3,8,8,8,9]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r9), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm8, %xmm4, %xmm7
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm7, %zmm10, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm6, %zmm23, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %ymm8, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm17[1,2,2,3,5,6,6,7]
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm8, %ymm19
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm17[0,0,2,1,4,4,6,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1,2],ymm8[3],ymm7[4,5],ymm8[6],ymm7[7,8,9,10],ymm8[11],ymm7[12,13],ymm8[14],ymm7[15]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm6, %zmm25, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 (%rax), %zmm25
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpermd %zmm6, %zmm29, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm7, %zmm28, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [6,7,3,3,7,7,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm27, %ymm28, %ymm6
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 96(%rax), %ymm7
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm30 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm5, %zmm30, %zmm12
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm9, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm10
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rcx), %xmm14
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm21[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0],ymm1[1],ymm9[2,3],ymm1[4],ymm9[5,6,7,8],ymm1[9],ymm9[10,11],ymm1[12],ymm9[13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,2,2,3,8,9,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm1, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm5[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = [0,0,2,1,8,8,9,11]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm22, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm11[4],xmm0[4],xmm11[5],xmm0[5],xmm11[6],xmm0[6],xmm11[7],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastq {{.*#+}} xmm24 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm11[0],xmm0[1],xmm11[2,3],xmm0[4],xmm11[5,6],xmm0[7]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm26, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm5[2,2,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} xmm29 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,1,1,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm20[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7,8,9,10],ymm2[11],ymm4[12,13],ymm2[14],ymm4[15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,1,3,3,8,8,9,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 100(%rax), %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 104(%rax), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [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]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm3, %zmm21, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r9), %xmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm18, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm0, %xmm8, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm11[0,1,2],ymm3[3],ymm11[4,5],ymm3[6],ymm11[7,8,9,10],ymm3[11],ymm11[12,13],ymm3[14],ymm11[15]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [6,7,3,3,7,7,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermd %ymm29, %ymm0, %ymm3
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 32(%rax), %ymm11
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm27
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm27
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rdi), %xmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm11, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm9[1,1,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm31, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm10
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm14, %xmm14
-; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm7, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm22, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm2, %xmm8, %xmm8
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,1,1,3,8,8,9,9]
-; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm0, %zmm2, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 36(%rax), %ymm0
-; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 40(%rax), %ymm7
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm6[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [2,2,2,3,8,8,8,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm9, %zmm17, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm23, %ymm28, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 32(%rax), %ymm20
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm9, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm31, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm11, %zmm30, %zmm20
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm16
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
+; AVX512F-ONLY-FAST-NEXT: vprold $16, %xmm5, %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7]
; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm19, %xmm13
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm13, %xmm5
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm15, %xmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm20, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm6[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
-; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0],xmm0[1],xmm9[2,3],xmm0[4],xmm9[5,6],xmm0[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm8 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm20 = mem[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm7 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm22 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpshufhw {{.*#+}} ymm6 = ymm0[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm6[3,3,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm6[2,2,2,2]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm30, %xmm15
-; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm15[0,2,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm6[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm5[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm17, %xmm5
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm1, %zmm22, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm1, %xmm14, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm8[0],xmm1[1],xmm8[2,3],xmm1[4],xmm8[5,6],xmm1[7]
; AVX512F-ONLY-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm6 = mem[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm9 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermpd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm31 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm30 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm30 = mem[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm29 = mem[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm16[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm13 = mem[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm10[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm11 = mem[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm12[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm15 = mem[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm14[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm28[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm26[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm25[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm23[2,1,3,2]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm21[2,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm8, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm7, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm10, %zmm8, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm20, %ymm7
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm10, %zmm3, %zmm4
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm14 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm4 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm5, %zmm3, %zmm10
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm10 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm21
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm22 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm20, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm3, %ymm2
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm20
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm7 # 64-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm19, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm17, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm18, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm16, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [0,0,1,1,8,8,10,9]
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm5, %zmm8, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm2, %zmm26, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm3, %xmm6, %xmm3
+; AVX512F-ONLY-FAST-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512F-ONLY-FAST-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm3, %zmm21, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm2
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm11, %ymm0
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm7
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512F-ONLY-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512F-ONLY-FAST-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm4 = mem[0,1,1,3,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7,8,9],ymm3[10],ymm4[11,12],ymm3[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm4 = 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]
+; AVX512F-ONLY-FAST-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # ymm5 = mem[1,1,1,1,5,5,5,5]
+; AVX512F-ONLY-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512F-ONLY-FAST-NEXT: vmovdqa 32(%rax), %ymm5
+; AVX512F-ONLY-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [14,21,0,0,15,22,0,15,14,21,0,0,15,22,0,15]
+; AVX512F-ONLY-FAST-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-FAST-NEXT: vpermi2d %zmm5, %zmm25, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,4,5,4,5,5,7]
+; AVX512F-ONLY-FAST-NEXT: vpermd %ymm5, %ymm8, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[16,17],zero,zero
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpandn %ymm8, %ymm9, %ymm8
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm6 # 64-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm9
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm6
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm30, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm28, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm25
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm28
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm1
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm0 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm2 = mem[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm2 = mem[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm5 = mem[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm5 = mem[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm23 = mem[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm26 = mem[0,0,1,1]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm6
+; AVX512F-ONLY-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm11, %ymm9
+; AVX512F-ONLY-FAST-NEXT: vpshufhw {{.*#+}} ymm11 = ymm15[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm11
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm9
+; AVX512F-ONLY-FAST-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm7[0,1,2,3],zmm9[4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: # xmm9 = mem[0,2,3,3,4,5,6,7]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
; AVX512F-ONLY-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
; AVX512F-ONLY-FAST-NEXT: # xmm11 = mem[0,2,3,3,4,5,6,7]
; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm12 = mem[0,2,2,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm13 = mem[0,1,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm15 = mem[2,1,3,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm16 = mem[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm17 = mem[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm18 = mem[0,0,1,3]
-; AVX512F-ONLY-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # ymm19 = mem[0,0,1,1]
-; AVX512F-ONLY-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
-; AVX512F-ONLY-FAST-NEXT: # xmm3 = mem[0,2,3,3,4,5,6,7]
-; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm12
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm8
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm12, %zmm0, %zmm27
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm5, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm26, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm2
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm17, %zmm0
-; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm19, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm3
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm5
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm24
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm14
-; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,3]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,3,2]
+; AVX512F-ONLY-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $184, %zmm7, %zmm8, %zmm14
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm1
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm0
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm1 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm0 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm3
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512F-ONLY-FAST-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]
+; AVX512F-ONLY-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm7 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm5 # 64-byte Folded Reload
+; AVX512F-ONLY-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
; AVX512F-ONLY-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm24, 256(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm27, 192(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm28, 128(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm9, 448(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 704(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm8, 640(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm25, 576(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm4, 512(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm22, 768(%rax)
-; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, 832(%rax)
-; AVX512F-ONLY-FAST-NEXT: addq $2264, %rsp # imm = 0x8D8
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm20, 192(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm13, (%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm18, 448(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm16, 704(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm12, 640(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512F-ONLY-FAST-NEXT: vmovaps %zmm2, 576(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm6, 384(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm1, 512(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm14, 832(%rax)
+; AVX512F-ONLY-FAST-NEXT: vmovdqa64 %zmm7, 768(%rax)
+; AVX512F-ONLY-FAST-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
+; AVX512DQ-SLOW-LABEL: store_i16_stride7_vf64:
+; AVX512DQ-SLOW: # %bb.0:
+; AVX512DQ-SLOW-NEXT: subq $2168, %rsp # imm = 0x878
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdx), %ymm6
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %ymm9
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm11, %ymm6, %ymm2
+; AVX512DQ-SLOW-NEXT: vporq %ymm1, %ymm2, %ymm18
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm14, %ymm9, %ymm1
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm13, %ymm7, %ymm2
+; AVX512DQ-SLOW-NEXT: vporq %ymm1, %ymm2, %ymm17
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa 64(%r9), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm3, %ymm8
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm21
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %ymm3
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm3, %ymm23
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rcx), %ymm10
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm10, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdx), %ymm4
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm11, %ymm4, %ymm2
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %ymm5
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm8, %ymm1, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm8, %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm2, %ymm2
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm11, %ymm2, %ymm2
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm14, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm24
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm8, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm8, %ymm19
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm8
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm11, %ymm2, %ymm1
+; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm14, %ymm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm13, %ymm11, %ymm13
+; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm13, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm13
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm13, %ymm14
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm12, %ymm0, %ymm15
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm12, %ymm20
+; AVX512DQ-SLOW-NEXT: vpor %ymm15, %ymm14, %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm0, %ymm14
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm13[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4],ymm14[5],ymm15[6,7,8,9],ymm14[10],ymm15[11,12],ymm14[13],ymm15[14,15]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm0[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm13[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm15[0,1],ymm12[2],ymm15[3,4],ymm12[5],ymm15[6,7,8,9],ymm12[10],ymm15[11,12],ymm12[13],ymm15[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-SLOW-NEXT: vpermi2q %zmm12, %zmm14, %zmm15
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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-SLOW-NEXT: # ymm14 = mem[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm16, %ymm15
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm14, %ymm15, %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm14, %ymm16
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7,8,9],ymm14[10],ymm12[11,12],ymm14[13],ymm12[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm15[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0],ymm12[1],ymm14[2,3],ymm12[4],ymm14[5,6,7,8],ymm12[9],ymm14[10,11],ymm12[12],ymm14[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0],ymm12[1],ymm14[2,3],ymm12[4],ymm14[5,6,7,8],ymm12[9],ymm14[10,11],ymm12[12],ymm14[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm12, %ymm25
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm9[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm14[0,1,2],ymm12[3],ymm14[4,5],ymm12[6],ymm14[7,8,9,10],ymm12[11],ymm14[12,13],ymm12[14],ymm14[15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm15[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm12[2],ymm6[3,4],ymm12[5],ymm6[6,7,8,9],ymm12[10],ymm6[11,12],ymm12[13],ymm6[14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm7 = ymm9[0,1],ymm7[2],ymm9[3,4],ymm7[5],ymm9[6,7,8,9],ymm7[10],ymm9[11,12],ymm7[13],ymm9[14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm18, %zmm6
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm17, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm7
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = ymm12[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm12[14,15],zero,zero,ymm12[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm12[16,17],zero,zero,ymm12[u,u],zero,zero
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, %ymm9, %ymm7, %ymm6
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %ymm15
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm20, %ymm14
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm14, %ymm15, %ymm14
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm14
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm7, %ymm6
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm7
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm15, %ymm6
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm6
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm14[0,1,2,3],zmm6[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm18 = [6,5,0,0,7,6,0,7,6,5,0,0,7,6,0,7]
+; AVX512DQ-SLOW-NEXT: # zmm18 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rax), %ymm6
+; AVX512DQ-SLOW-NEXT: vpermd %zmm6, %zmm18, %zmm12
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm6[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [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-SLOW-NEXT: vpshufb %ymm15, %ymm6, %ymm6
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm14, %ymm9, %ymm14
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm6, %zmm6
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 72(%rax), %ymm6
+; AVX512DQ-SLOW-NEXT: vpandnq %ymm6, %ymm17, %ymm14
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rax), %ymm6
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm6, %ymm12
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm14, %zmm12
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 8(%rax), %ymm12
+; AVX512DQ-SLOW-NEXT: vpandnq %ymm12, %ymm17, %ymm12
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rax), %ymm7
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm7, %ymm14
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm7, %ymm22
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm12
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm1[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm14[2],ymm12[3,4],ymm14[5],ymm12[6,7,8,9],ymm14[10],ymm12[11,12],ymm14[13],ymm12[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rax), %ymm12
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm14, %ymm9, %ymm9
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm15, %ymm12, %ymm14
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm14, %zmm9
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm13[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm9[3],ymm0[4,5],ymm9[6],ymm0[7,8,9,10],ymm9[11],ymm0[12,13],ymm9[14],ymm0[15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm16, %ymm7
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm7, %ymm8, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm9[2],ymm0[3,4],ymm9[5],ymm0[6,7,8,9],ymm9[10],ymm0[11,12],ymm9[13],ymm0[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm8[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = 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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm1[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm0, %ymm31
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpermd %zmm12, %zmm18, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm5[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm5[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm10[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %ymm7, %ymm10, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm21, %ymm9
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm21, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm9[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm23[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-SLOW-NEXT: vpermt2q %zmm0, %zmm7, %zmm1
+; AVX512DQ-SLOW-NEXT: vbroadcasti32x8 {{.*#+}} zmm15 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512DQ-SLOW-NEXT: # zmm15 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermd 64(%rax), %zmm15, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm14, %zmm0
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm5[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm23[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm9[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm4 = 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-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [4,5,4,5,4,5,6,7,16,17,16,17,16,17,17,19]
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm3[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm11, %zmm4
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm6[2,3,3,3,6,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 96(%rax), %ymm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm1, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rsi), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdi), %xmm5
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm4, %xmm6
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm6, %xmm23
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rcx), %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqa 96(%rdx), %xmm5
+; AVX512DQ-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm13 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm4, %xmm6
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm9 = xmm5[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0],xmm6[1],xmm9[2,3],xmm6[4],xmm9[5,6],xmm6[7]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm30 = <16,18,19,19,19,19,u,u,0,1,0,1,2,3,2,3>
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm3, %zmm30, %zmm2
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 100(%rax), %ymm3
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 104(%rax), %ymm4
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [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-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm28, %zmm3
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rcx), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdx), %xmm3
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3],xmm2[4],xmm3[5,6],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm5 = <16,16,17,17,17,17,u,u,0,1,0,1,2,3,2,3>
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm3, %zmm5, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rdi), %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%rsi), %xmm4
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm6, %xmm6
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm9, %xmm27
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm4, %xmm4
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4],xmm4[5],xmm3[6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm4
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r9), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa 64(%r8), %xmm4
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <0,1,0,1,0,1,1,3,16,18,19,19,19,19,u,u>
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm4, %zmm6, %zmm2
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 64(%rax), %ymm4
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 68(%rax), %ymm9
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm4, %zmm4
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm9, %zmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm26
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm2 = xmm10[0],xmm2[1],xmm10[2,3],xmm2[4],xmm10[5,6],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm10 = xmm12[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm10, %zmm5, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm10
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm12, %xmm12
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm4, %xmm21
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm10, %xmm10
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm10[2],xmm5[3,4],xmm10[5],xmm5[6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm12, %zmm4
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm3, %zmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm3
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm3, %zmm6, %zmm2
+; AVX512DQ-SLOW-NEXT: vpbroadcastd (%rax), %ymm3
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm9, %zmm3
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm24, %ymm10
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm10[1,2,3,3,4,5,6,7,9,10,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm19[1,1,1,1,5,5,5,5]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm20
+; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vprold $16, %ymm3, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm9[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm5 = ymm3[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm9[0,0,2,1,4,4,6,5]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm6[3],ymm5[4,5],ymm6[6],ymm5[7,8,9,10],ymm6[11],ymm5[12,13],ymm6[14],ymm5[15]
+; AVX512DQ-SLOW-NEXT: vpermt2q %zmm2, %zmm7, %zmm5
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm10[0,1,2,3,5,4,6,7,8,9,10,11,13,12,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm19[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3],ymm2[4],ymm6[5,6,7,8],ymm2[9],ymm6[10,11],ymm2[12],ymm6[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-SLOW-NEXT: vpermd (%rax), %zmm15, %zmm2
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm14, %zmm2
+; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm6[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,0,0,4,4,4,4]
+; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm17
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[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-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm6[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2,3],ymm5[4],ymm2[5,6,7,8],ymm5[9],ymm2[10,11],ymm5[12],ymm2[13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm10[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm9[3,3,3,3,7,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm3[0,1,2,3,5,6,7,7,8,9,10,11,13,14,15,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,3,6,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0,1],ymm2[2],ymm5[3,4],ymm2[5],ymm5[6,7,8,9],ymm2[10],ymm5[11,12],ymm2[13],ymm5[14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm6
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm14
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm5[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm12, %zmm11, %zmm2
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm11 = ymm22[2,3,3,3,6,7,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 32(%rax), %ymm12
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm22
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm22
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512DQ-SLOW-NEXT: vprold $16, %xmm2, %xmm12
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm1[1,1,2,3]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm3 = xmm15[0,1],xmm12[2],xmm15[3,4],xmm12[5],xmm15[6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm8, %xmm2
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm3
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm23, %xmm4
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512DQ-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm24 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm19 = mem[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm8 = mem[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,5,7,6]
+; AVX512DQ-SLOW-NEXT: vpermt2d %zmm1, %zmm30, %zmm0
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512DQ-SLOW-NEXT: vpbroadcastd 40(%rax), %ymm5
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm30
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm28, %zmm30
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm0
+; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm0, %xmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm14
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[1,1,2,2]
+; AVX512DQ-SLOW-NEXT: vpblendw {{.*#+}} xmm1 = xmm15[0],xmm5[1],xmm15[2,3],xmm5[4],xmm15[5,6],xmm5[7]
+; AVX512DQ-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm14[4],xmm0[4],xmm14[5],xmm0[5],xmm14[6],xmm0[6],xmm14[7],xmm0[7]
+; AVX512DQ-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm0[3,3,3,3]
+; AVX512DQ-SLOW-NEXT: vpshufd $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm23 = mem[1,2,2,3,5,6,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm13 = ymm4[2,1,2,3,6,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm13 = ymm13[0,0,3,3,4,5,6,7,8,8,11,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm13[2,2,2,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm2[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm29, %xmm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm2[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm27, %xmm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm9 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm3[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm26, %xmm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm3[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm5[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm5 = mem[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm10 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $250, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm11 = mem[2,2,3,3]
+; AVX512DQ-SLOW-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm0 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm29 = mem[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm31[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm26 = mem[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm21 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm15 = mem[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm4 = mem[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm12 = mem[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm18[2,2,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm24, %zmm24
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm25, %zmm0
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpternlogq $184, %zmm24, %zmm27, %zmm0
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm23[2,1,3,2]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm8
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm28
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm13 = [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-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm8 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm23 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm23[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13, %zmm23 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm13 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm13, %zmm24, %zmm7
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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-SLOW-NEXT: vpternlogq $226, %zmm8, %zmm9, %zmm7
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm24, %zmm8
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm23, %zmm9, %zmm8
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm3
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm18
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm18
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm19 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm2[0,1,2,3],zmm5[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm23 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vextracti64x4 $1, %zmm0, %ymm0
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-SLOW-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm2, %ymm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm28[0,1,2,3],zmm0[4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm25
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm28 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm13 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm9 # 64-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm4, %zmm1
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [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-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm1
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm20, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm17, %zmm5
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm3, %zmm5
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm20
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm21 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm21
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm26, %zmm31, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm27, %zmm1
+; AVX512DQ-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm0 = mem[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm3 = mem[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,1,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm3 = mem[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm5 = mem[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
+; AVX512DQ-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm5 = mem[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm6 = mem[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm26 = mem[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm10 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm11 = mem[0,2,2,3]
+; AVX512DQ-SLOW-NEXT: vpshuflw $180, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm12 = mem[0,1,3,2,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm14 = mem[2,1,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm15 = mem[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm16 = mem[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw $230, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm2 = mem[2,1,2,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,5,4]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,3]
+; AVX512DQ-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # ymm17 = mem[0,0,1,1]
+; AVX512DQ-SLOW-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
+; AVX512DQ-SLOW-NEXT: # xmm4 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm27, %zmm3
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm12[0,1,1,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm11, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm11
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm27, %zmm11
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm12
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm11, %zmm0, %zmm22
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm26, %zmm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm24, %zmm3
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm16, %zmm0
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm17, %zmm2
+; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm24, %zmm2
+; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm4
+; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm30
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm23
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm28, %zmm0
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm23, 320(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm30, 256(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm22, 192(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm21, 128(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm9, (%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm13, 448(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, 704(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm12, 640(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm20, 576(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, 512(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, 384(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm19, 768(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm25, 832(%rax)
+; AVX512DQ-SLOW-NEXT: addq $2168, %rsp # imm = 0x878
+; AVX512DQ-SLOW-NEXT: vzeroupper
+; AVX512DQ-SLOW-NEXT: retq
+;
; AVX512DQ-FAST-LABEL: store_i16_stride7_vf64:
; AVX512DQ-FAST: # %bb.0:
-; AVX512DQ-FAST-NEXT: subq $2264, %rsp # imm = 0x8D8
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm4
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm2
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm3
-; AVX512DQ-FAST-NEXT: vporq %ymm2, %ymm3, %ymm16
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm4, %ymm2
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm11, %ymm10, %ymm3
-; AVX512DQ-FAST-NEXT: vporq %ymm2, %ymm3, %ymm17
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [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-FAST-NEXT: vmovdqa 64(%r9), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: subq $1432, %rsp # imm = 0x598
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [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-FAST-NEXT: vpshufb %ymm11, %ymm4, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %ymm5
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512DQ-FAST-NEXT: vporq %ymm0, %ymm1, %ymm16
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [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-FAST-NEXT: vpshufb %ymm15, %ymm6, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <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-FAST-NEXT: vpshufb %ymm8, %ymm7, %ymm1
+; AVX512DQ-FAST-NEXT: vporq %ymm0, %ymm1, %ymm19
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [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-FAST-NEXT: vmovdqa 64(%r9), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u>
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vporq %ymm2, %ymm3, %ymm21
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm6
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm6, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vporq %ymm2, %ymm3, %ymm20
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vpor %ymm2, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm15, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm5, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm1, %ymm1
+; AVX512DQ-FAST-NEXT: vporq %ymm0, %ymm1, %ymm22
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %ymm1
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm1, %ymm1
; AVX512DQ-FAST-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm2
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm11
-; AVX512DQ-FAST-NEXT: vpor %ymm0, %ymm11, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm14
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm14, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm12
-; AVX512DQ-FAST-NEXT: vporq %ymm12, %ymm11, %ymm22
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm11 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vprold $16, %ymm1, %ymm12
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm14[1,2,2,3,5,6,6,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm11, %zmm12, %zmm13
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <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-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm10[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <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-FAST-NEXT: vpshufb %ymm12, %ymm4, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm12, %ymm18
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} 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-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm11
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm19
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm8[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm13
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm30
+; AVX512DQ-FAST-NEXT: vpor %ymm13, %ymm9, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm13
+; AVX512DQ-FAST-NEXT: vporq %ymm9, %ymm13, %ymm17
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm13
+; AVX512DQ-FAST-NEXT: vpor %ymm9, %ymm13, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm13
+; AVX512DQ-FAST-NEXT: vpor %ymm9, %ymm13, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %ymm3
+; AVX512DQ-FAST-NEXT: vpshufb %ymm11, %ymm3, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm11
+; AVX512DQ-FAST-NEXT: vpor %ymm9, %ymm11, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm15, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm8
+; AVX512DQ-FAST-NEXT: vpor %ymm9, %ymm8, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %ymm11
+; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm11, %ymm8
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %ymm15
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm15, %ymm9
+; AVX512DQ-FAST-NEXT: vporq %ymm9, %ymm8, %ymm18
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[0,1,1,3,4,5,5,7]
; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm0, %ymm31
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-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-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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-FAST-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm12, %ymm31
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-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-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm16, %zmm8
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm17, %zmm9
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm19, %zmm9
; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm9
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <5,u,u,u,6,u,u,6>
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm10
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <u,4,u,u,u,5,u,u>
-; AVX512DQ-FAST-NEXT: vpermd %ymm10, %ymm8, %ymm8
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[14,15],zero,zero,ymm10[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm10[16,17],zero,zero,ymm10[u,u],zero,zero
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm23 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $248, %ymm23, %ymm9, %ymm10
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm0, %ymm11
-; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10, %ymm11
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %ymm8
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm12 = ymm8[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm8[14,15],zero,zero,ymm8[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm8[16,17],zero,zero,ymm8[u,u],zero,zero
+; AVX512DQ-FAST-NEXT: vpternlogq $248, %ymm13, %ymm9, %ymm12
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm13, %ymm14
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %ymm13
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm13, %ymm10
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm12, %ymm10
; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm9
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm8
-; AVX512DQ-FAST-NEXT: vprold $16, %ymm0, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <u,4,u,u,u,5,u,u>
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm12, %ymm12
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm12
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm13, %ymm9
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm8, %ymm9
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm8
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm11[0,1,2,3],zmm8[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm12, %ymm9
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm9
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm9 = zmm10[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 64(%rax), %zmm26
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm9 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
-; AVX512DQ-FAST-NEXT: # zmm9 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rax), %ymm8
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm26, %zmm8, %zmm9
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rax), %ymm9
+; AVX512DQ-FAST-NEXT: vpermd %ymm9, %ymm10, %ymm10
+; AVX512DQ-FAST-NEXT: vpandn %ymm10, %ymm14, %ymm10
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm13, %ymm9, %ymm12
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <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-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm10
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm12, %ymm19
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = <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-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm12, %ymm23
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm6[1],ymm7[2,3],ymm6[4],ymm7[5,6,7,8],ymm6[9],ymm7[10,11],ymm6[12],ymm7[13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = [0,2,2,3,10,9,11,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm25, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = <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-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm10
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm6, %ymm26
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-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-FAST-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-FAST-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm6, %ymm4, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm6, %ymm29
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm20, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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-FAST-NEXT: vpternlogq $226, %zmm7, %zmm5, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = <5,u,u,u,6,u,u,6>
+; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm4, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa64 64(%rax), %zmm28
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm4 = [30,5,0,0,31,6,0,31,30,5,0,0,31,6,0,31]
+; AVX512DQ-FAST-NEXT: # zmm4 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm28, %zmm9, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm7 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpbroadcastd 72(%rax), %ymm4
+; AVX512DQ-FAST-NEXT: vpandn %ymm4, %ymm7, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm7, %ymm16
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rax), %ymm5
+; AVX512DQ-FAST-NEXT: vmovdqa %ymm13, %ymm14
+; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm5, %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm5, %ymm27
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm4, %zmm9
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm4
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm12
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm5, %xmm21
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm7, %xmm7
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm13, %xmm8
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [0,0,1,1,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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-FAST-NEXT: vpternlogq $248, %zmm4, %zmm7, %zmm9
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm8, %ymm24, %ymm9
-; AVX512DQ-FAST-NEXT: vpandnq %ymm9, %ymm23, %ymm9
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [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-FAST-NEXT: vpshufb %ymm12, %ymm8, %ymm8
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpbroadcastd 72(%rax), %ymm8
-; AVX512DQ-FAST-NEXT: vpandn %ymm8, %ymm9, %ymm8
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm9, %ymm17
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rax), %ymm11
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm11, %ymm9
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm16
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm10
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r9), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%r8), %xmm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,0,1,1,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm10, %zmm21, %zmm13
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [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-FAST-NEXT: vpternlogq $248, %zmm27, %zmm13, %zmm16
+; AVX512DQ-FAST-NEXT: vpbroadcastd 8(%rax), %ymm7
+; AVX512DQ-FAST-NEXT: vpandnq %ymm7, %ymm16, %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm5
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm9
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm16
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm5
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, (%rsp) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm14 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm14, %xmm14
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm13, %zmm14
+; AVX512DQ-FAST-NEXT: vpternlogq $248, %zmm4, %zmm14, %zmm16
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpbroadcastd 8(%rax), %ymm10
-; AVX512DQ-FAST-NEXT: vpandnq %ymm10, %ymm17, %ymm17
-; AVX512DQ-FAST-NEXT: vmovdqa (%rax), %ymm10
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm10, %ymm13
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm10, %ymm29
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm17, %zmm10
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm0, %zmm20
-; AVX512DQ-FAST-NEXT: vmovdqa (%r9), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%r8), %xmm7
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm7[4],xmm4[4],xmm7[5],xmm4[5],xmm7[6],xmm4[6],xmm7[7],xmm4[7]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm13, %xmm13
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm20, %zmm21, %zmm13
-; AVX512DQ-FAST-NEXT: vpternlogq $248, %zmm27, %zmm13, %zmm10
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm13 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm5[0,1,1,3,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm4
-; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm2, %ymm10
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm3[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1],ymm13[2],ymm10[3,4],ymm13[5],ymm10[6,7,8,9],ymm13[10],ymm10[11,12],ymm13[13],ymm10[14,15]
+; AVX512DQ-FAST-NEXT: vpbroadcastd {{.*#+}} ymm9 = [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-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm4
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm7[3],ymm4[4,5],ymm7[6],ymm4[7,8,9,10],ymm7[11],ymm4[12,13],ymm7[14],ymm4[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = <0,1,u,3,10,10,11,11>
+; AVX512DQ-FAST-NEXT: vpermi2q %zmm4, %zmm18, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-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-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm10 = ymm14[0,0,2,1,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm10[3],ymm0[4,5],ymm10[6],ymm0[7,8,9,10],ymm10[11],ymm0[12,13],ymm10[14],ymm0[15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = <0,1,u,3,10,10,11,11>
-; AVX512DQ-FAST-NEXT: vpermi2q %zmm0, %zmm22, %zmm10
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm0
-; AVX512DQ-FAST-NEXT: vpermd %ymm0, %ymm24, %ymm10
-; AVX512DQ-FAST-NEXT: vpandnq %ymm10, %ymm23, %ymm10
-; AVX512DQ-FAST-NEXT: vpshufb %ymm12, %ymm0, %ymm12
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm28, %ymm13
-; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm2, %ymm10
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm12 = ymm3[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = 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-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm18, %ymm14
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm3[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm25, %ymm3
-; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm19, %ymm14
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm0, %ymm4
; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm19, %ymm10
-; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm15, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa64 (%rax), %zmm16
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm1 = [14,21,0,0,15,22,0,15,14,21,0,0,15,22,0,15]
-; AVX512DQ-FAST-NEXT: # zmm1 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermi2d %zmm0, %zmm16, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm13, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm2[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm20
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm7[0,1,2],ymm4[3],ymm7[4,5],ymm4[6],ymm7[7,8,9,10],ymm4[11],ymm7[12,13],ymm4[14],ymm7[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm23, %ymm5
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm7
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm13[0],ymm7[1],ymm13[2,3],ymm7[4],ymm13[5,6,7,8],ymm7[9],ymm13[10,11],ymm7[12],ymm13[13,14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm25, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm26, %ymm8
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm3, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm29, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm6
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm13 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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-FAST-NEXT: vpermt2q %zmm4, %zmm20, %zmm0
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm7, %zmm10, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm4 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-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-FAST-NEXT: vprold $16, %ymm15, %ymm6
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm11[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = 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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm17, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm2[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm4
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm30[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqa64 %ymm31, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm6
; AVX512DQ-FAST-NEXT: vmovdqa %ymm2, %ymm5
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vmovdqa %ymm4, %ymm13
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vprold $16, %ymm3, %ymm0
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm1 = ymm4[1,2,2,3,5,6,6,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %ymm7, %ymm3, %ymm1
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[0,0,2,1,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm12, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm26, %zmm0
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm1 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512DQ-FAST-NEXT: # zmm1 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %zmm0, %zmm1, %zmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm1, %zmm0
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm13[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3],ymm0[4],ymm2[5,6,7,8],ymm0[9],ymm2[10,11],ymm0[12],ymm2[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm6, %ymm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5],ymm0[6],ymm2[7,8,9,10],ymm0[11],ymm2[12,13],ymm0[14],ymm2[15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm0 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = 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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm5
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm3
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [6,7,3,3,7,7,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm11, %ymm3, %ymm3
-; AVX512DQ-FAST-NEXT: vpbroadcastd 96(%rax), %ymm6
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm3, %zmm3
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm3
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm3
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm2, %xmm6
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm3[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2],xmm7[3,4],xmm6[5],xmm7[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm6, %xmm24
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm2, %xmm20
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm6
-; AVX512DQ-FAST-NEXT: vpbroadcastq {{.*#+}} xmm13 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm2, %xmm3
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm3 = xmm7[0],xmm3[1],xmm7[2,3],xmm3[4],xmm7[5,6],xmm3[7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm31
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm4
-; AVX512DQ-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm4, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm4, %xmm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm4, %zmm0
-; AVX512DQ-FAST-NEXT: vpbroadcastd 100(%rax), %ymm2
-; AVX512DQ-FAST-NEXT: vpbroadcastd 104(%rax), %ymm4
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm30[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1],ymm7[2],ymm6[3,4],ymm7[5],ymm6[6,7,8,9],ymm7[10],ymm6[11,12],ymm7[13],ymm6[14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm20, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %ymm1, %ymm0, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm29, %ymm31
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm1[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm7[2],ymm4[3,4],ymm7[5],ymm4[6,7,8,9],ymm7[10],ymm4[11,12],ymm7[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm7 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm11[0,1],ymm7[2],ymm11[3,4],ymm7[5],ymm11[6,7,8,9],ymm7[10],ymm11[11,12],ymm7[13],ymm11[14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm17, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm23 = [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-FAST-NEXT: vpternlogq $226, %zmm6, %zmm23, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm15, %ymm4
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm19[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = 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-FAST-NEXT: vpshufb %ymm9, %ymm15, %ymm6
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm19[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm11[3],ymm6[4,5],ymm11[6],ymm6[7,8,9,10],ymm11[11],ymm6[12,13],ymm11[14],ymm6[15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm25 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm25, %zmm6
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm27, %zmm28, %zmm4
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm29 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512DQ-FAST-NEXT: # zmm29 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermd %zmm4, %zmm29, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = [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-FAST-NEXT: vpternlogq $184, %zmm6, %zmm28, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm2
; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rcx), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm15, %xmm30
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm13, %xmm10
-; AVX512DQ-FAST-NEXT: vpshufb %xmm13, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2,3],xmm0[4],xmm2[5,6],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FAST-NEXT: vpshufb %ymm8, %ymm0, %ymm3
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm10 = ymm2[0],ymm3[1],ymm2[2,3],ymm3[4],ymm2[5,6,7,8],ymm3[9],ymm2[10,11],ymm3[12],ymm2[13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rcx), %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdx), %xmm11
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,2,2,3,8,9,9,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm0, %zmm10
+; AVX512DQ-FAST-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm0 = ymm30[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm5 = 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-FAST-NEXT: vmovdqa 64(%rcx), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdx), %xmm7
+; AVX512DQ-FAST-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, %xmm8
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0],xmm0[1],xmm3[2,3],xmm0[4],xmm3[5,6],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm1[4],xmm7[5],xmm1[5],xmm7[6],xmm1[6],xmm7[7],xmm1[7]
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm3, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm3, %zmm13
; AVX512DQ-FAST-NEXT: vmovdqa 64(%rdi), %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm2
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm13 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm13, %xmm19
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm2, %xmm2
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2],xmm0[3,4],xmm2[5],xmm0[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm5, %xmm9
-; AVX512DQ-FAST-NEXT: vpshufb %xmm5, %xmm0, %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm14, %xmm11
-; AVX512DQ-FAST-NEXT: vpshufb %xmm14, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,0,1,8,9,9,11]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm5, %zmm0
-; AVX512DQ-FAST-NEXT: vpbroadcastd 64(%rax), %ymm2
-; AVX512DQ-FAST-NEXT: vpbroadcastd 68(%rax), %ymm6
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm2, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm7
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm0, %xmm17
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm10, %xmm6, %xmm6
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3],xmm6[4],xmm7[5,6],xmm6[7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512DQ-FAST-NEXT: vmovdqa 64(%rsi), %xmm3
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm3, %xmm4
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2],xmm6[3,4],xmm4[5],xmm6[6,7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm14
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm26, %zmm14
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm14
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm21, %xmm0
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm0[0],xmm12[1],xmm0[1],xmm12[2],xmm0[2],xmm12[3],xmm0[3]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm3, %xmm12
+; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [0,0,0,1,8,9,9,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm21, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa (%rcx), %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm7
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm3[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3],xmm7[4],xmm8[5,6],xmm7[7]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm8 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm2, %xmm24
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm13, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FAST-NEXT: vmovdqa (%rsi), %xmm8
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm15 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm8, %xmm8
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm8[2],xmm6[3,4],xmm8[5],xmm6[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm6, %ymm28
-; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm6 = xmm6[0],mem[0],xmm6[1],mem[1],xmm6[2],mem[2],xmm6[3],mem[3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm6, %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm9, %xmm22
-; AVX512DQ-FAST-NEXT: vpshufb %xmm11, %xmm6, %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm11, %xmm18
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm8, %zmm5, %zmm6
-; AVX512DQ-FAST-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512DQ-FAST-NEXT: vpbroadcastd 4(%rax), %ymm8
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm6, %zmm2, %zmm5
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[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-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[1,1,1,1,5,5,5,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm6[2],ymm2[3,4],ymm6[5],ymm2[6,7,8,9],ymm6[10],ymm2[11,12],ymm6[13],ymm2[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm26
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm10[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-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm11[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0],ymm2[1],ymm6[2,3],ymm2[4],ymm6[5,6,7,8],ymm2[9],ymm6[10,11],ymm2[12],ymm6[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm25
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vprold $16, %ymm5, %ymm2
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm8, %xmm3
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm2[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm3[2],xmm7[3,4],xmm3[5],xmm7[6,7]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm9, %xmm2
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm26, %zmm3
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm1, %zmm4, %zmm3
+; AVX512DQ-FAST-NEXT: vpbroadcastd 64(%rax), %ymm1
+; AVX512DQ-FAST-NEXT: vpbroadcastd 68(%rax), %ymm4
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm18
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm18
+; AVX512DQ-FAST-NEXT: vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm4
+; AVX512DQ-FAST-NEXT: vpshufb %xmm6, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqa %xmm6, %xmm8
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm21, %zmm0
+; AVX512DQ-FAST-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512DQ-FAST-NEXT: vpbroadcastd 4(%rax), %ymm6
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm13
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm13
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rsi), %xmm9
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%rdi), %xmm0
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm5
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm5
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm19[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
+; AVX512DQ-FAST-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-FAST-NEXT: vpternlogq $184, %zmm14, %zmm2, %zmm18
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm2, %zmm13
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-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-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-FAST-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FAST-NEXT: vpshufb %ymm3, %ymm4, %ymm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm4, %ymm16
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm20[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm6 = 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-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm3, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm31, %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb %ymm2, %ymm15, %ymm2
; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm7[1,2,2,3,5,6,6,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0,1],ymm2[2],ymm6[3,4],ymm2[5],ymm6[6,7,8,9],ymm2[10],ymm6[11,12],ymm2[13],ymm6[14,15]
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm6 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm7[0,0,2,1,4,4,6,5]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1,2],ymm8[3],ymm6[4,5],ymm8[6],ymm6[7,8,9,10],ymm8[11],ymm6[12,13],ymm8[14],ymm6[15]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm12, %zmm8
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm9[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[0,1,1,3,4,5,5,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm6[0,1],ymm2[2],ymm6[3,4],ymm2[5],ymm6[6,7,8,9],ymm2[10],ymm6[11,12],ymm2[13],ymm6[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm6 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm6[2],ymm2[3,4],ymm6[5],ymm2[6,7,8,9],ymm6[10],ymm2[11,12],ymm6[13],ymm2[14,15]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm2, %ymm21
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm16, %zmm2
-; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512DQ-FAST-NEXT: # zmm6 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %zmm2, %zmm6, %zmm2
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm8, %zmm1, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm4, %xmm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm1, %ymm16
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm12
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm6
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm9 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm7[2,2,2,2,6,6,6,6]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4],ymm3[5],ymm2[6,7,8,9],ymm3[10],ymm2[11,12],ymm3[13],ymm2[14,15]
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm15[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm7, %ymm21
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm14 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7,8,9],ymm3[10],ymm4[11,12],ymm3[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm17, %zmm14
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r9), %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa 96(%r8), %xmm2
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm8, %xmm4, %xmm7
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm7, %zmm10, %zmm1
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm6, %zmm23, %zmm14
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vprold $16, %ymm8, %ymm6
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm7 = ymm17[1,2,2,3,5,6,6,7]
+; AVX512DQ-FAST-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-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm8, %ymm19
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm8 = ymm17[0,0,2,1,4,4,6,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1,2],ymm8[3],ymm7[4,5],ymm8[6],ymm7[7,8,9,10],ymm8[11],ymm7[12,13],ymm8[14],ymm7[15]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm6, %zmm25, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa64 (%rax), %zmm25
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm6
+; AVX512DQ-FAST-NEXT: vpermd %zmm6, %zmm29, %zmm10
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm7, %zmm28, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} ymm28 = [6,7,3,3,7,7,6,7]
+; AVX512DQ-FAST-NEXT: vpermd %ymm27, %ymm28, %ymm6
+; AVX512DQ-FAST-NEXT: vpbroadcastd 96(%rax), %ymm7
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [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-FAST-NEXT: vpternlogq $184, %zmm1, %zmm31, %zmm12
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm30 = [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-FAST-NEXT: vpternlogq $184, %zmm5, %zmm30, %zmm12
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm9[4],xmm0[4],xmm9[5],xmm0[5],xmm9[6],xmm0[6],xmm9[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm9, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm7 = xmm0[0,1],xmm1[2],xmm0[3,4],xmm1[5],xmm0[6,7]
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm10
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rcx), %xmm14
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm15[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm9 = ymm21[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0],ymm1[1],ymm9[2,3],ymm1[4],ymm9[5,6,7,8],ymm1[9],ymm9[10,11],ymm1[12],ymm9[13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,2,2,3,8,9,9,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm1, %zmm6
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm5[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm22 = [0,0,2,1,8,8,9,11]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm22, %zmm7
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm11[4],xmm0[4],xmm11[5],xmm0[5],xmm11[6],xmm0[6],xmm11[7],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpbroadcastq {{.*#+}} xmm24 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm11[0],xmm0[1],xmm11[2,3],xmm0[4],xmm11[5,6],xmm0[7]
+; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm5, %zmm26, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm26 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm1
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm15[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-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm5[2,2,2,2]
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} xmm29 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm27 = [0,1,1,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm16, %ymm2
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} ymm4 = ymm20[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm2[3],ymm4[4,5],ymm2[6],ymm4[7,8,9,10],ymm2[11],ymm4[12,13],ymm2[14],ymm4[15]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm2
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm5
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
+; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm4, %xmm4
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm4, %zmm0, %zmm11
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm11
+; AVX512DQ-FAST-NEXT: vpbroadcastd 100(%rax), %ymm0
+; AVX512DQ-FAST-NEXT: vpbroadcastd 104(%rax), %ymm4
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm16
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm21 = [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-FAST-NEXT: vpternlogq $184, %zmm3, %zmm21, %zmm16
+; AVX512DQ-FAST-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[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-FAST-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[3,3,3,3,7,7,7,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4],ymm3[5],ymm0[6,7,8,9],ymm3[10],ymm0[11,12],ymm3[13],ymm0[14,15]
; AVX512DQ-FAST-NEXT: vmovdqa 32(%r9), %xmm4
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm18, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufb %xmm0, %xmm8, %xmm3
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm3, %zmm27, %zmm9
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm3 = 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-FAST-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm0 = ymm11[0,1,2],ymm3[3],ymm11[4,5],ymm3[6],ymm11[7,8,9,10],ymm3[11],ymm11[12,13],ymm3[14],ymm11[15]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [6,7,3,3,7,7,6,7]
-; AVX512DQ-FAST-NEXT: vpermd %ymm29, %ymm0, %ymm3
-; AVX512DQ-FAST-NEXT: vpbroadcastd 32(%rax), %ymm11
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm27
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm27
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rdi), %xmm9
-; AVX512DQ-FAST-NEXT: vmovdqa 32(%rsi), %xmm11
-; AVX512DQ-FAST-NEXT: vprold $16, %xmm11, %xmm0
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm7 = xmm9[1,1,2,3]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm7[0,1],xmm0[2],xmm7[3,4],xmm0[5],xmm7[6,7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm31, %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm13, %xmm10
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm14, %xmm14
-; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm7, %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm22, %xmm2
-; AVX512DQ-FAST-NEXT: vpshufb %xmm2, %xmm8, %xmm8
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,1,1,3,8,8,9,9]
-; AVX512DQ-FAST-NEXT: vpermt2q %zmm0, %zmm2, %zmm8
-; AVX512DQ-FAST-NEXT: vpbroadcastd 36(%rax), %ymm0
-; AVX512DQ-FAST-NEXT: vpbroadcastd 40(%rax), %ymm7
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm24
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm24
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%r8), %xmm3
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm6[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm9, %zmm17, %zmm0
+; AVX512DQ-FAST-NEXT: vpermd %ymm23, %ymm28, %ymm9
+; AVX512DQ-FAST-NEXT: vpbroadcastd 32(%rax), %ymm20
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm9, %zmm20
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm31, %zmm20
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm11, %zmm30, %zmm20
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm16
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm2[4],xmm5[5],xmm2[5],xmm5[6],xmm2[6],xmm5[7],xmm2[7]
+; AVX512DQ-FAST-NEXT: vprold $16, %xmm5, %xmm5
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7]
; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} xmm9 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm19, %xmm13
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm13, %xmm5
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm15, %xmm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm20, %xmm3
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm3, %xmm3
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm0 = xmm6[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm9 = xmm1[1,1,2,2]
-; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0],xmm0[1],xmm9[2,3],xmm0[4],xmm9[5,6],xmm0[7]
-; AVX512DQ-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm8 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm20 = mem[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm7 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm22 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
-; AVX512DQ-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpshufhw {{.*#+}} ymm6 = ymm0[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm6[3,3,3,3]
-; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm6[2,2,2,2]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm30, %xmm15
-; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm6 = xmm15[0,2,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm6[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm5[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm17, %xmm5
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm1, %zmm22, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm24, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm1, %xmm14, %xmm1
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
+; AVX512DQ-FAST-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} xmm1 = xmm8[0],xmm1[1],xmm8[2,3],xmm1[4],xmm8[5,6],xmm1[7]
; AVX512DQ-FAST-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm6 = mem[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm9 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermpd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermpd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FAST-NEXT: vpermpd $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vmovups %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm31 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm31 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $182, {{[-0-9]+}}(%r{{[sb]}}p), %ymm30 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm30 = mem[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %ymm29 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm29 = mem[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm19 = ymm16[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm13 = mem[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm17 = ymm10[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm11 = mem[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm18 = ymm12[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm15 = mem[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm16 = ymm14[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm28[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm26[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm26 = ymm25[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm25 = ymm23[2,1,3,2]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm21[2,2,2,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm8, %zmm10
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm22, %zmm7, %zmm20
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [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-FAST-NEXT: vpternlogq $226, %zmm10, %zmm8, %zmm20
-; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm20, %ymm7
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [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-FAST-NEXT: vpternlogq $226, %zmm10, %zmm3, %zmm4
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm4 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5, %zmm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm10 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm5, %zmm3, %zmm10
-; AVX512DQ-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm10 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm21
-; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [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-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm22 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm22 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm14 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm20, %ymm2
-; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm3, %ymm2
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm1 = [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-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm0, %zmm1, %zmm20
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm7 # 64-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm19, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm17, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [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-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm18, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm16, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [0,0,1,1,8,8,10,9]
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm5, %zmm8, %zmm1
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm2, %zmm26, %zmm1
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FAST-NEXT: vmovdqa64 %xmm29, %xmm3
+; AVX512DQ-FAST-NEXT: vpshufb %xmm3, %xmm6, %xmm3
+; AVX512DQ-FAST-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
+; AVX512DQ-FAST-NEXT: vpbroadcastd 36(%rax), %ymm2
+; AVX512DQ-FAST-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm21, %zmm2
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm2
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm11, %ymm0
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm7
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
+; AVX512DQ-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm1
+; AVX512DQ-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm0
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FAST-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm4 = mem[0,1,1,3,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4],ymm3[5],ymm4[6,7,8,9],ymm3[10],ymm4[11,12],ymm3[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm4 = 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-FAST-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # ymm5 = mem[1,1,1,1,5,5,5,5]
+; AVX512DQ-FAST-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
+; AVX512DQ-FAST-NEXT: vmovdqa 32(%rax), %ymm5
+; AVX512DQ-FAST-NEXT: vbroadcasti32x8 {{.*#+}} zmm6 = [14,21,0,0,15,22,0,15,14,21,0,0,15,22,0,15]
+; AVX512DQ-FAST-NEXT: # zmm6 = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermi2d %zmm5, %zmm25, %zmm6
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FAST-NEXT: vpermd %ymm5, %ymm8, %ymm8
+; AVX512DQ-FAST-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm5[16,17],zero,zero
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-FAST-NEXT: vpandn %ymm8, %ymm9, %ymm8
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [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-FAST-NEXT: vpternlogq $184, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm6 # 64-byte Folded Reload
; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm9
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm6
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm0, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm30, %zmm1
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [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-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm1
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm26, %zmm28, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm2, %zmm5
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm25 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm1, %zmm0, %zmm25
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm5, %zmm0, %zmm28
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm1
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm0 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm2 = mem[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
-; AVX512DQ-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm2 = mem[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm5 = mem[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
-; AVX512DQ-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm5 = mem[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm23 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm23 = mem[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm26 = mem[0,0,1,1]
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm6
+; AVX512DQ-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm9
+; AVX512DQ-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FAST-NEXT: vpternlogd $226, 124(%r8){1to8}, %ymm11, %ymm9
+; AVX512DQ-FAST-NEXT: vpshufhw {{.*#+}} ymm11 = ymm15[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[3,3,3,3]
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm9, %ymm11
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm9
+; AVX512DQ-FAST-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm7[0,1,2,3],zmm9[4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload
+; AVX512DQ-FAST-NEXT: # xmm9 = mem[0,2,3,3,4,5,6,7]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,1,3]
; AVX512DQ-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
; AVX512DQ-FAST-NEXT: # xmm11 = mem[0,2,3,3,4,5,6,7]
; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm12 = mem[0,2,2,3]
-; AVX512DQ-FAST-NEXT: vpermq $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm13 = mem[0,1,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm15 = mem[2,1,3,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm16 = mem[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpermq $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm17 = mem[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpermq $208, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm18 = mem[0,0,1,3]
-; AVX512DQ-FAST-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # ymm19 = mem[0,0,1,1]
-; AVX512DQ-FAST-NEXT: vpshuflw $248, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
-; AVX512DQ-FAST-NEXT: # xmm3 = mem[0,2,3,3,4,5,6,7]
-; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,2,1]
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm16, %zmm15, %zmm12
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm8, %zmm12
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm8
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm12, %zmm0, %zmm27
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm5, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm26, %zmm2
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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-FAST-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm2
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm18, %zmm17, %zmm0
-; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm19, %zmm3
-; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm3
-; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm2, %zmm0, %zmm5
-; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm3, %zmm0, %zmm24
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm14
-; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,1,3]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,3,2]
+; AVX512DQ-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $184, %zmm7, %zmm8, %zmm14
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm1
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $226, %zmm7, %zmm26, %zmm0
+; AVX512DQ-FAST-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm1 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm0 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm3
+; AVX512DQ-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm7, %zmm4
+; AVX512DQ-FAST-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
+; AVX512DQ-FAST-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-FAST-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm7 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm5 # 64-byte Folded Reload
+; AVX512DQ-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm5
; AVX512DQ-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 320(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm24, 256(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm27, 192(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm28, 128(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 64(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, (%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm9, 448(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 704(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm8, 640(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm25, 576(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm4, 512(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm22, 768(%rax)
-; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, 832(%rax)
-; AVX512DQ-FAST-NEXT: addq $2264, %rsp # imm = 0x8D8
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm5, 320(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm2, 256(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm20, 192(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm13, (%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm18, 448(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm16, 704(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm12, 640(%rax)
+; AVX512DQ-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FAST-NEXT: vmovaps %zmm2, 576(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm6, 384(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm1, 512(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm14, 832(%rax)
+; AVX512DQ-FAST-NEXT: vmovdqa64 %zmm7, 768(%rax)
+; AVX512DQ-FAST-NEXT: addq $1432, %rsp # imm = 0x598
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
@@ -13504,10 +14471,8 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-SLOW: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512BW-ONLY-SLOW: {{.*}}
-; AVX512DQ-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512DQBW-SLOW: {{.*}}
-; AVX512F-ONLY-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-4.ll
index 15fa50557067273..c0200ee6f6ebcf5 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-4.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i64-stride-4.ll
@@ -159,26 +159,44 @@ define void @store_i64_stride4_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
-; AVX512-LABEL: store_i64_stride4_vf4:
-; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512-NEXT: vmovdqa (%rdx), %ymm2
-; AVX512-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm0[0,1],ymm2[0,1]
-; AVX512-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm1[0,1],ymm3[0,1]
-; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
-; AVX512-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm3[2,3]
-; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
-; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
-; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
-; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm1
-; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512-NEXT: vmovdqa64 %zmm0, 64(%r8)
-; AVX512-NEXT: vmovdqa64 %zmm1, (%r8)
-; AVX512-NEXT: vzeroupper
-; AVX512-NEXT: retq
+; AVX512-SLOW-LABEL: store_i64_stride4_vf4:
+; AVX512-SLOW: # %bb.0:
+; AVX512-SLOW-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512-SLOW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-SLOW-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512-SLOW-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512-SLOW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm0[0,1],ymm2[0,1]
+; AVX512-SLOW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm1[0,1],ymm3[0,1]
+; AVX512-SLOW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
+; AVX512-SLOW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm3[2,3]
+; AVX512-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
+; AVX512-SLOW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
+; AVX512-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
+; AVX512-SLOW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
+; AVX512-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm1
+; AVX512-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512-SLOW-NEXT: vmovdqa64 %zmm0, 64(%r8)
+; AVX512-SLOW-NEXT: vmovdqa64 %zmm1, (%r8)
+; AVX512-SLOW-NEXT: vzeroupper
+; AVX512-SLOW-NEXT: retq
+;
+; AVX512-FAST-LABEL: store_i64_stride4_vf4:
+; AVX512-FAST: # %bb.0:
+; AVX512-FAST-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512-FAST-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-FAST-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512-FAST-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512-FAST-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm0[0,1],ymm2[0,1]
+; AVX512-FAST-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm1[0,1],ymm3[0,1]
+; AVX512-FAST-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
+; AVX512-FAST-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],ymm3[2,3]
+; AVX512-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,8,2,10,1,9,3,11]
+; AVX512-FAST-NEXT: vpermt2q %zmm5, %zmm2, %zmm4
+; AVX512-FAST-NEXT: vpermt2q %zmm1, %zmm2, %zmm0
+; AVX512-FAST-NEXT: vmovdqa64 %zmm0, 64(%r8)
+; AVX512-FAST-NEXT: vmovdqa64 %zmm4, (%r8)
+; AVX512-FAST-NEXT: vzeroupper
+; AVX512-FAST-NEXT: retq
%in.vec0 = load <4 x i64>, ptr %in.vecptr0, align 64
%in.vec1 = load <4 x i64>, ptr %in.vecptr1, align 64
%in.vec2 = load <4 x i64>, ptr %in.vecptr2, align 64
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 f75d96b70dcbd03..1f21eee473194bd 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
@@ -902,13 +902,13 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-SLOW-LABEL: store_i8_stride5_vf16:
; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm1
+; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm2
; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm3
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm1
; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm5
-; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm6
+; AVX512F-SLOW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm6
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,7],zero,ymm6[u,u,u,8],zero,ymm6[u,u,u,9],zero,ymm6[u,u,u],zero,ymm6[26,u,u,u],zero,ymm6[27,u,u,u],zero,ymm6[28,u,u]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm6[2,3,0,1]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u],zero,ymm8[7,u,u,u],zero,ymm8[8,u,u,u],zero,ymm8[9,u,u,u,26],zero,ymm8[u,u,u,27],zero,ymm8[u,u,u,28],zero,ymm8[u,u]
@@ -925,20 +925,17 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[0,8],zero,zero,ymm6[u,1,9],zero,zero,ymm6[u,2,10],zero,zero,ymm6[u,3,19],zero,zero,ymm6[u,28,20],zero,zero,ymm6[u,29,21],zero,zero,ymm6[u,30,22]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
; AVX512F-SLOW-NEXT: vporq %zmm7, %zmm5, %zmm5
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm0[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm0[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm6 = <u,0,0,0,0,u,1,1,1,1,u,2,2,2,2,u>
+; AVX512F-SLOW-NEXT: vpermd %zmm1, %zmm6, %zmm6
; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[u],zero,zero,xmm3[10,11,u],zero,zero,xmm3[12,13,u],zero,zero,xmm3[14,15,u]
-; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,10,11],zero,zero,xmm1[u,12,13],zero,zero,xmm1[u,14,15],zero,zero,xmm1[u]
-; AVX512F-SLOW-NEXT: vpor %xmm3, %xmm1, %xmm1
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[1,2,3,4],zero,xmm1[6,7,8,9],zero,xmm1[11,12,13,14],zero
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[12],zero,zero,zero,zero,xmm0[13],zero,zero,zero,zero,xmm0[14],zero,zero,zero,zero,xmm0[15]
-; AVX512F-SLOW-NEXT: vpor %xmm0, %xmm1, %xmm0
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,10,11],zero,zero,xmm0[u,12,13],zero,zero,xmm0[u,14,15],zero,zero,xmm0[u]
+; AVX512F-SLOW-NEXT: vpor %xmm3, %xmm0, %xmm0
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = zero,xmm0[1,2,3,4],zero,xmm0[6,7,8,9],zero,xmm0[11,12,13,14],zero
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[12],zero,zero,zero,zero,xmm1[13],zero,zero,zero,zero,xmm1[14],zero,zero,zero,zero,xmm1[15]
+; AVX512F-SLOW-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX512F-SLOW-NEXT: vmovdqa %xmm0, 64(%r9)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm6, (%r9)
; AVX512F-SLOW-NEXT: vzeroupper
@@ -1009,11 +1006,8 @@ define void @store_i8_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,ymm5[0,8],zero,zero,zero,ymm5[1,9],zero,zero,zero,ymm5[2,10],zero,zero,zero,ymm5[19,27],zero,zero,zero,ymm5[20,28],zero,zero,zero,ymm5[21,29],zero,zero,zero
; AVX512BW-SLOW-NEXT: vpor %ymm6, %ymm5, %ymm5
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[0,0,1,1]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm4[1,1,2,2]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,1,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512BW-SLOW-NEXT: vpermd %zmm4, %zmm6, %zmm6
; AVX512BW-SLOW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm6, %zmm5 {%k1}
@@ -1890,11 +1884,8 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-SLOW-NEXT: vpternlogq $226, %ymm8, %ymm7, %ymm5
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm6
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm6
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <u,0,0,0,0,u,1,1,1,1,u,2,2,2,2,u>
+; AVX512F-SLOW-NEXT: vpermd %zmm6, %zmm8, %zmm6
; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm6
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u],zero,ymm4[13,u,u,u],zero,ymm4[14,u,u,u],zero,ymm4[15,u,u,u],zero,ymm4[16,u,u,u],zero,ymm4[17,u,u,u],zero,ymm4[18,u,u]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm3[u,u,u,13],zero,ymm3[u,u,u,14],zero,ymm3[u,u,u,15],zero,ymm3[u,u,u,16],zero,ymm3[u,u,u,17],zero,ymm3[u,u,u,18],zero,ymm3[u,u]
@@ -2055,11 +2046,8 @@ define void @store_i8_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm6
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[0,0,1,1]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,1,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
+; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512BW-SLOW-NEXT: vpermd %zmm6, %zmm7, %zmm6
; AVX512BW-SLOW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm6, %zmm3 {%k1}
@@ -3898,192 +3886,178 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-SLOW-LABEL: store_i8_stride5_vf64:
; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm4
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm3
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm4, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm0
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm6 = <12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19>
-; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm2, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm7 = <12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19>
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm2, %ymm1
; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm1
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm8 = <8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11>
; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm1, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm1, %xmm18
-; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm3
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm1, %xmm29
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm4
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128>
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm3, %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm3, %xmm17
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm4, %xmm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm4, %xmm30
; AVX512F-SLOW-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm9
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm9, %ymm3
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm9, %ymm4
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm11
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = <u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u>
; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm11, %ymm10
-; AVX512F-SLOW-NEXT: vpor %ymm3, %ymm10, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpor %ymm4, %ymm10, %ymm4
+; AVX512F-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm13
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = <128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12>
-; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm13, %xmm3
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm13, %xmm4
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm5, %xmm25
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm10
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = <6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128>
; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm10, %xmm12
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm5, %xmm26
-; AVX512F-SLOW-NEXT: vporq %xmm3, %xmm12, %xmm21
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm22
+; AVX512F-SLOW-NEXT: vporq %xmm4, %xmm12, %xmm31
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm22
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm2[26],zero,ymm2[28],zero,zero,ymm2[27],zero,ymm2[29],zero,ymm2[31],zero,zero,ymm2[30],zero
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[21],zero,zero,ymm2[20],zero,ymm2[22],zero,ymm2[24],zero,zero,ymm2[23],zero,ymm2[25],zero,zero
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm23
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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]
-; AVX512F-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm11, %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm30
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm9, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm31
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,ymm11[26],zero,ymm11[28],zero,ymm11[30],zero,zero,ymm11[29],zero,ymm11[31],zero,zero
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm9[19],zero,ymm9[21],zero,zero,ymm9[20],zero,ymm9[22],zero,ymm9[24],zero,zero,ymm9[23],zero
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm24
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm7, %ymm0
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm12
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm12, %ymm1
-; AVX512F-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm19
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm12
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm12, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm6
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm6, %ymm1
+; AVX512F-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm20
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm5
-; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm5, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm15, %ymm5, %ymm2
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm4, %ymm1
-; AVX512F-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm20
+; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm4, %ymm3
+; AVX512F-SLOW-NEXT: vporq %ymm2, %ymm3, %ymm21
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm0, %xmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm0, %xmm16
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm2, %xmm6
-; AVX512F-SLOW-NEXT: vporq %xmm1, %xmm6, %xmm28
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm0, %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm0, %xmm17
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm8
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm8, %xmm3
+; AVX512F-SLOW-NEXT: vporq %xmm2, %xmm3, %xmm27
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm25, %xmm0
-; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm6, %xmm0
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm1, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm1, %xmm19
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm2
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm26, %xmm1
-; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm3, %xmm15
-; AVX512F-SLOW-NEXT: vporq %xmm0, %xmm15, %xmm29
-; AVX512F-SLOW-NEXT: vmovdqa64 32(%r8), %xmm25
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm25[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,1,1]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
-; AVX512F-SLOW-NEXT: vpandnq %ymm15, %ymm26, %ymm15
-; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm2, %xmm15
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm16
+; AVX512F-SLOW-NEXT: vporq %xmm0, %xmm15, %xmm28
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = mem[1,1,2,2]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,1,1]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255]
+; AVX512F-SLOW-NEXT: vpandnq %ymm0, %ymm25, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm15
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm0, %ymm14
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm15, %zmm26
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm15
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm15[0,2,1,1,4,6,5,5]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,3,2]
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
-; AVX512F-SLOW-NEXT: vpandnq %ymm14, %ymm27, %ymm14
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm15, %ymm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm1, %zmm27
-; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm1 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm9, %ymm9
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm31, %ymm8
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm7, %ymm14
-; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm1
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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]
-; AVX512F-SLOW-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm11, %ymm11
-; AVX512F-SLOW-NEXT: vpshufb %ymm7, %ymm12, %ymm7
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm30, %ymm8
-; AVX512F-SLOW-NEXT: vpshufb %ymm8, %ymm12, %ymm12
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm15, %ymm14
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm26
+; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm0, %ymm1
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255]
+; AVX512F-SLOW-NEXT: vpandn %ymm0, %ymm14, %ymm0
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm25
+; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm9, %ymm9
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm14 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm12[19],zero,ymm12[21],zero,zero,ymm12[20],zero,ymm12[22],zero,ymm12[24],zero,zero,ymm12[23],zero
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm12, %ymm12
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512F-SLOW-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm11, %ymm11
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm6, %ymm2
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[27],zero,zero,ymm6[26],zero,ymm6[28],zero,ymm6[30],zero,zero,ymm6[29],zero,ymm6[31],zero,zero
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3],xmm13[4],xmm10[4],xmm13[5],xmm10[5],xmm13[6],xmm10[6],xmm13[7],xmm10[7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm9[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm11[2,2,3,3]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm13 = ymm5[3,u,5,u,u,4,u,6,u,8,u,u,7,u,9,u,19,u,21,u,u,20,u,22,u,24,u,u,23,u,25,u]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,ymm4[27],zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30],zero
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm18, %xmm11
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm17, %xmm8
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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]
-; AVX512F-SLOW-NEXT: vmovdqa64 %xmm16, %xmm8
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3],xmm8[4],xmm2[4],xmm8[5],xmm2[5],xmm8[6],xmm2[6],xmm8[7],xmm2[7]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm8 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
-; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm11, %xmm11
-; AVX512F-SLOW-NEXT: vpshufb %xmm8, %xmm2, %xmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm0[0,2,1,1,4,6,5,5]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm5[11,u,u,10,u,12,u,u,u,u,13,u,15,u,u,14,27,u,u,26,u,28,u,u,u,u,29,u,31,u,u,30]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm4[21],zero,zero,ymm4[20],zero,ymm4[22],zero,ymm4[24],zero,zero,ymm4[23],zero,ymm4[25],zero,zero
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm4[26],zero,ymm4[28],zero,zero,ymm4[27],zero,ymm4[29],zero,ymm4[31],zero,zero,ymm4[30],zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm29, %xmm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
+; AVX512F-SLOW-NEXT: vmovdqa64 (%r8), %zmm29
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <4,u,5,5,5,5,u,6,6,6,6,u,7,7,7,7>
+; AVX512F-SLOW-NEXT: vpermd %zmm15, %zmm1, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = <6,6,6,u,7,7,7,7,u,16,16,16,16,u,17,17>
+; AVX512F-SLOW-NEXT: vpermi2d %zmm15, %zmm29, %zmm1
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm9[2,2,3,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm11[2,2,3,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,3,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,3]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm28, %zmm2, %zmm28
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm10, %xmm10
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,3,3]
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm17, %xmm9
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm3, %xmm3
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm8, %xmm8
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm10, %xmm10
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,1]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[2,2,3,3,6,6,7,7]
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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]
-; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm3, %xmm2
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm29, %zmm2, %zmm2
-; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: # ymm3 = mem[0,0,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm21[0,0,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm16 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm3, %zmm16, %zmm6
-; AVX512F-SLOW-NEXT: vpor %ymm7, %ymm14, %ymm3
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm19, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm7 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
-; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm7, %ymm13, %ymm9
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm20, %zmm9
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm3, %zmm16, %zmm9
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm3 = zmm22[2,2,3,3,6,6,7,7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm27, %zmm8, %zmm8
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm19, %xmm9
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm16, %xmm11
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3],xmm9[4],xmm11[4],xmm9[5],xmm11[5],xmm9[6],xmm11[6],xmm9[7],xmm11[7]
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm9, %xmm9
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm28, %zmm9, %zmm9
+; AVX512F-SLOW-NEXT: vpermq $80, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm15 = mem[0,0,1,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm15 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm31[0,0,1,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm16, %zmm16 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm17 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm15, %zmm17, %zmm16
+; AVX512F-SLOW-NEXT: vpor %ymm2, %ymm14, %ymm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm20, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [18374966859431608575,18374966859431608575,18446463693966278400,18446463693966278400]
+; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm14, %ymm13, %ymm0
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm21, %zmm0
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm2, %zmm17, %zmm0
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm22[2,2,3,3,6,6,7,7]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm13 = zmm23[2,2,3,3,6,6,7,7]
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm13
-; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm7, %ymm1, %ymm12
-; AVX512F-SLOW-NEXT: vpandq %ymm7, %ymm30, %ymm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm31, %zmm1
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm3 = zmm24[2,2,3,3,6,6,7,7]
-; AVX512F-SLOW-NEXT: vporq %zmm3, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm13, %zmm3, %zmm1
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm7
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm4
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm7, %zmm3, %zmm4
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm26
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm8[2,3,3,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,3,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm27
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm15[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm25[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm1
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm3
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm28[0,0,1,1,4,4,5,5]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,0,1,1,4,4,5,5]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm2
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[0,0,1,1]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[1,1,2,2]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,1,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm3
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm27, 64(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm3, (%r9)
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm13
+; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm14, %ymm12, %ymm6
+; AVX512F-SLOW-NEXT: vpandq %ymm14, %ymm30, %ymm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm18, %zmm2
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm24[2,2,3,3,6,6,7,7]
+; AVX512F-SLOW-NEXT: vporq %zmm11, %zmm2, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm13, %zmm11, %zmm2
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm6, %zmm6
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7, %ymm5
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm5, %zmm3
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm11, %zmm3
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm26
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm4
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm25
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm1
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm9[0,0,1,1,4,4,5,5]
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm0 = <u,0,0,0,0,u,1,1,1,1,u,2,2,2,2,u>
+; AVX512F-SLOW-NEXT: vpermd %zmm29, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm25, 64(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, (%r9)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, 256(%r9)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, 256(%r9)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm26, 192(%r9)
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
@@ -4269,28 +4243,29 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512BW-ONLY-SLOW-LABEL: store_i8_stride5_vf64:
; AVX512BW-ONLY-SLOW: # %bb.0:
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r8), %zmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm8, %ymm0, %ymm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: movl $693250386, %eax # imm = 0x29522952
; AVX512BW-ONLY-SLOW-NEXT: kmovd %eax, %k1
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm4, %ymm2 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm5
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm3, %ymm2
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm4[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm4, %xmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm10
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm15 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm4, %ymm2
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm5[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,3,3,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: movl $1251232404, %eax # imm = 0x4A944A94
; AVX512BW-ONLY-SLOW-NEXT: kmovd %eax, %k5
@@ -4298,81 +4273,76 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm13
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm14
-; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm11 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm11, %xmm9, %xmm9
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm9 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k4
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm5, %zmm2 {%k4}
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k4}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r8), %ymm16
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,3,3,6,6,7,7]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm15
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[0,0,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [6,6,6,6,7,7,7,7,16,16,16,16,16,16,17,17]
+; AVX512BW-ONLY-SLOW-NEXT: vpermi2d %zmm16, %zmm3, %zmm10
; AVX512BW-ONLY-SLOW-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k2
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm9, %zmm2 {%k2}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm23
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm23, %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm10, %ymm23, %ymm17
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm24
-; AVX512BW-ONLY-SLOW-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]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm10, %ymm24, %ymm19
-; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm18, %ymm19, %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm24, %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm17, %ymm18, %ymm17
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm20 = <128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm20, %xmm12, %xmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm22 = <6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm22, %xmm6, %xmm6
; AVX512BW-ONLY-SLOW-NEXT: vpor %xmm6, %xmm12, %xmm6
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm6, %zmm6
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm19 = <8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm19, %xmm14, %xmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm21 = <128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm21, %xmm13, %xmm13
; AVX512BW-ONLY-SLOW-NEXT: vpor %xmm12, %xmm13, %xmm12
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm12[0,0,1,1]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm12[0,0,1,1]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = <3,3,3,u,4,4,4,4>
-; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm25, %ymm12, %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm25, %ymm12, %ymm17
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm26
-; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm14 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512BW-ONLY-SLOW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512BW-ONLY-SLOW-NEXT: kmovd %eax, %k3
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm26, %ymm18 {%k3}
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm14, %zmm14
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm26, %ymm17 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm13, %zmm13
; AVX512BW-ONLY-SLOW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k2
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm14, %zmm6 {%k2}
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = <3,3,3,3,u,4,4,4>
-; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm16, %ymm14, %ymm18
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,2,2]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm15, %zmm15
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm13, %zmm6 {%k2}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm13 = <3,3,3,3,u,4,4,4>
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm16, %ymm13, %ymm17
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm18 = mem[1,1,2,2]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,1,1,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm18, %zmm17
; AVX512BW-ONLY-SLOW-NEXT: movabsq $4760450083537948804, %rax # imm = 0x4210842108421084
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k6
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm15, %zmm6 {%k6}
-; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
-; AVX512BW-ONLY-SLOW-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm26, %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm17, %zmm6 {%k6}
+; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
+; AVX512BW-ONLY-SLOW-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm26, %ymm18
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm18[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm18 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm18, %ymm25, %ymm28
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm27, %ymm28, %ymm27
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm26, %ymm17
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm26, %ymm15
; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm25 = ymm25[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,2,3,3,6,6,7,7]
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm25, %ymm17 {%k5}
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,3,3]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm27, %zmm25
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm23, %ymm26
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm25, %ymm15 {%k5}
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm27, %zmm15
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm23, %ymm26
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm27 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX512BW-ONLY-SLOW-NEXT: # ymm27 = mem[0,1,2,3,0,1,2,3]
@@ -4385,72 +4355,66 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm23, %ymm8 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm26, %zmm8
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm25, %zmm8 {%k4}
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm23 = ymm16[0,2,1,1,4,6,5,5]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,3,2]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm16 = ymm16[2,2,3,3,6,6,7,7]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,3,3]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm23, %zmm16
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdx), %xmm23
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k4}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %zmm16, %zmm15, %zmm15
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdx), %xmm16
; AVX512BW-ONLY-SLOW-NEXT: movabsq $-8925843906633654008, %rax # imm = 0x8421084210842108
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm16, %zmm8 {%k1}
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rcx), %xmm16
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm20, %xmm16, %xmm20
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm22, %xmm23, %xmm22
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm15
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm20, %xmm15, %xmm20
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm22, %xmm16, %xmm22
; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm20, %xmm22, %xmm20
-; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm16[0],xmm23[0],xmm16[1],xmm23[1],xmm16[2],xmm23[2],xmm16[3],xmm23[3],xmm16[4],xmm23[4],xmm16[5],xmm23[5],xmm16[6],xmm23[6],xmm16[7],xmm23[7]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm16, %xmm7
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rsi), %xmm16
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm15, %xmm7
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm15
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm20, %zmm7, %zmm7
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %xmm20
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm19, %xmm20, %xmm19
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm21, %xmm16, %xmm21
-; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm19, %xmm21, %xmm19
-; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm20[0],xmm16[0],xmm20[1],xmm16[1],xmm20[2],xmm16[2],xmm20[3],xmm16[3],xmm20[4],xmm16[4],xmm20[5],xmm16[5],xmm20[6],xmm16[6],xmm20[7],xmm16[7]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm11, %xmm16, %xmm11
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %xmm16
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm19, %xmm16, %xmm19
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm21, %xmm15, %xmm20
+; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm19, %xmm20, %xmm19
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm9, %xmm15, %xmm9
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm9, %zmm9
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,0,1,1,4,4,5,5]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm11, %zmm11
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,0,1,1,4,4,5,5]
; AVX512BW-ONLY-SLOW-NEXT: movabsq $-4165393823095705204, %rax # imm = 0xC6318C6318C6318C
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm7, %zmm11 {%k1}
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm16 = xmm7[0,0,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,0,0,1]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,1,1]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm16, %zmm7
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm7, %zmm9 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %zmm3, %zmm7, %zmm3
; AVX512BW-ONLY-SLOW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm7, %zmm11 {%k1}
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm1, %ymm7
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm3, %zmm9 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm1, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm0, %ymm7
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm0, %ymm16
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,3,3]
-; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm7, %ymm16, %ymm7
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm1, %ymm1
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm11, %ymm0, %ymm0
; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm4, %ymm12, %ymm1
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm1 {%k3}
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm3
-; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm18, %ymm4, %ymm4
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm5, %ymm12, %ymm1
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm4, %ymm1 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm17, %ymm4, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm18, %ymm5, %ymm4
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
-; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm5, %ymm14, %ymm0
-; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[0,2,1,1,4,6,5,5]
-; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,3,2]
-; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm0
+; AVX512BW-ONLY-SLOW-NEXT: vpermd %ymm0, %ymm13, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, 64(%r9)
-; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, (%r9)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, (%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 256(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, 192(%r9)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm2, 128(%r9)
@@ -4614,28 +4578,29 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQBW-SLOW-LABEL: store_i8_stride5_vf64:
; AVX512DQBW-SLOW: # %bb.0:
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r8), %zmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rcx), %ymm0
; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12,9,14,11,0,13,10,15,12]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm8, %ymm0, %ymm2
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm4 = ymm1[0,1,2,3,5,6,7,6,8,9,10,11,13,14,15,14]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,2,3,3,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: movl $693250386, %eax # imm = 0x29522952
; AVX512DQBW-SLOW-NEXT: kmovd %eax, %k1
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm3, %ymm2 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm4, %ymm2 {%k1}
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm6
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rcx), %xmm12
-; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm6[0],xmm12[1],xmm6[1],xmm12[2],xmm6[2],xmm12[3],xmm6[3],xmm12[4],xmm6[4],xmm12[5],xmm6[5],xmm12[6],xmm6[6],xmm12[7],xmm6[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8>
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm5
-; AVX512DQBW-SLOW-NEXT: vmovdqa (%rsi), %ymm3
-; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm3, %ymm2
-; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdi), %ymm4
-; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm4[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm7, %xmm4, %xmm4
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm10
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rsi), %ymm4
+; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm15 = [11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14,11,0,13,10,15,12,0,14]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm15, %ymm4, %ymm2
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdi), %ymm5
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm9 = ymm5[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,2,3,3,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: movl $1251232404, %eax # imm = 0x4A944A94
; AVX512DQBW-SLOW-NEXT: kmovd %eax, %k5
@@ -4643,81 +4608,76 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rsi), %xmm13
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm14
-; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm11 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm11, %xmm9, %xmm9
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm9 = <0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm9, %xmm11, %xmm11
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
; AVX512DQBW-SLOW-NEXT: movabsq $1785168781326730801, %rax # imm = 0x18C6318C6318C631
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k4
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm5, %zmm2 {%k4}
-; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %ymm5
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k4}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm16
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,3,3,6,6,7,7]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,3,3]
-; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%r8), %xmm15
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm15[0,0,1,1]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [6,6,6,6,7,7,7,7,16,16,16,16,16,16,17,17]
+; AVX512DQBW-SLOW-NEXT: vpermi2d %zmm16, %zmm3, %zmm10
; AVX512DQBW-SLOW-NEXT: movabsq $2380225041768974402, %rax # imm = 0x2108421084210842
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k2
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm9, %zmm2 {%k2}
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm10, %zmm2 {%k2}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm23
-; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm9, %ymm23, %ymm18
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,12,13,128,128,128,128,14,128,128,128,14,15,128,128,128,128,16,128,128,128,16,17,128,128,128,128,18,128,128,128]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm10, %ymm23, %ymm17
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm24
-; AVX512DQBW-SLOW-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]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm10, %ymm24, %ymm19
-; AVX512DQBW-SLOW-NEXT: vporq %ymm18, %ymm19, %ymm18
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm11, %ymm24, %ymm18
+; AVX512DQBW-SLOW-NEXT: vporq %ymm17, %ymm18, %ymm17
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm20 = <128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12>
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm20, %xmm12, %xmm12
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm22 = <6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128>
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm22, %xmm6, %xmm6
; AVX512DQBW-SLOW-NEXT: vpor %xmm6, %xmm12, %xmm6
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm6, %zmm6
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm19 = <8,128,u,7,128,9,128,u,128,u,10,128,12,128,u,11>
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm19, %xmm14, %xmm12
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm21 = <128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128>
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm21, %xmm13, %xmm13
; AVX512DQBW-SLOW-NEXT: vpor %xmm12, %xmm13, %xmm12
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm12[0,0,1,1]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm12[0,0,1,1]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm25
; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = <3,3,3,u,4,4,4,4>
-; AVX512DQBW-SLOW-NEXT: vpermd %ymm25, %ymm12, %ymm18
+; AVX512DQBW-SLOW-NEXT: vpermd %ymm25, %ymm12, %ymm17
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm26
-; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
+; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm14 = [0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14,0,0,13,2,15,0,1,14]
; AVX512DQBW-SLOW-NEXT: movl $138547332, %eax # imm = 0x8421084
; AVX512DQBW-SLOW-NEXT: kmovd %eax, %k3
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm13, %ymm26, %ymm18 {%k3}
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm14, %zmm14
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm14, %ymm26, %ymm17 {%k3}
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm13, %zmm13
; AVX512DQBW-SLOW-NEXT: movabsq $-8330787646191410408, %rax # imm = 0x8C6318C6318C6318
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k2
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm14, %zmm6 {%k2}
-; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = <3,3,3,3,u,4,4,4>
-; AVX512DQBW-SLOW-NEXT: vpermd %ymm16, %ymm14, %ymm18
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[1,1,2,2]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,1,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm15, %zmm15
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm13, %zmm6 {%k2}
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm13 = <3,3,3,3,u,4,4,4>
+; AVX512DQBW-SLOW-NEXT: vpermd %ymm16, %ymm13, %ymm17
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm18 = mem[1,1,2,2]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm18[0,1,1,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm18, %zmm17
; AVX512DQBW-SLOW-NEXT: movabsq $4760450083537948804, %rax # imm = 0x4210842108421084
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k6
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm15, %zmm6 {%k6}
-; AVX512DQBW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
-; AVX512DQBW-SLOW-NEXT: # ymm15 = mem[0,1,0,1]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm15, %ymm26, %ymm18
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm17, %zmm6 {%k6}
+; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm17 = [19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128]
+; AVX512DQBW-SLOW-NEXT: # ymm17 = mem[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm26, %ymm18
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm18[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm18 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm18, %ymm25, %ymm28
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vporq %ymm27, %ymm28, %ymm27
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm26, %ymm17
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm15, %ymm26, %ymm15
; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm25 = ymm25[0,1,2,3,6,5,6,7,8,9,10,11,14,13,14,15]
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm25 = ymm25[2,2,3,3,6,6,7,7]
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm25, %ymm17 {%k5}
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm17 = ymm17[2,2,3,3]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm27, %zmm25
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm17 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm23, %ymm26
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm25, %ymm15 {%k5}
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,3]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm27, %zmm15
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm23, %ymm26
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm27 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
; AVX512DQBW-SLOW-NEXT: # ymm27 = mem[0,1,0,1]
@@ -4730,72 +4690,66 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm23, %ymm8 {%k1}
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm26, %zmm8
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm25, %zmm8 {%k4}
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm23 = ymm16[0,2,1,1,4,6,5,5]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,3,2]
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm16 = ymm16[2,2,3,3,6,6,7,7]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,3,3]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm23, %zmm16
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdx), %xmm23
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k4}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm15 = [4,6,5,5,5,5,4,6,6,6,6,6,7,7,7,7]
+; AVX512DQBW-SLOW-NEXT: vpermd %zmm16, %zmm15, %zmm15
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdx), %xmm16
; AVX512DQBW-SLOW-NEXT: movabsq $-8925843906633654008, %rax # imm = 0x8421084210842108
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm16, %zmm8 {%k1}
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rcx), %xmm16
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm20, %xmm16, %xmm20
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm22, %xmm23, %xmm22
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm15, %zmm8 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rcx), %xmm15
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm20, %xmm15, %xmm20
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm22, %xmm16, %xmm22
; AVX512DQBW-SLOW-NEXT: vporq %xmm20, %xmm22, %xmm20
-; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm16[0],xmm23[0],xmm16[1],xmm23[1],xmm16[2],xmm23[2],xmm16[3],xmm23[3],xmm16[4],xmm23[4],xmm16[5],xmm23[5],xmm16[6],xmm23[6],xmm16[7],xmm23[7]
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm7, %xmm16, %xmm7
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rsi), %xmm16
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm15[0],xmm16[0],xmm15[1],xmm16[1],xmm15[2],xmm16[2],xmm15[3],xmm16[3],xmm15[4],xmm16[4],xmm15[5],xmm16[5],xmm15[6],xmm16[6],xmm15[7],xmm16[7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm7, %xmm15, %xmm7
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rsi), %xmm15
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm20, %zmm7, %zmm7
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %xmm20
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm19, %xmm20, %xmm19
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm21, %xmm16, %xmm21
-; AVX512DQBW-SLOW-NEXT: vporq %xmm19, %xmm21, %xmm19
-; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm16 = xmm20[0],xmm16[0],xmm20[1],xmm16[1],xmm20[2],xmm16[2],xmm20[3],xmm16[3],xmm20[4],xmm16[4],xmm20[5],xmm16[5],xmm20[6],xmm16[6],xmm20[7],xmm16[7]
-; AVX512DQBW-SLOW-NEXT: vpshufb %xmm11, %xmm16, %xmm11
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %xmm16
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm19, %xmm16, %xmm19
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm21, %xmm15, %xmm20
+; AVX512DQBW-SLOW-NEXT: vporq %xmm19, %xmm20, %xmm19
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm9, %xmm15, %xmm9
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm9, %zmm9
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,0,1,1,4,4,5,5]
-; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm11, %zmm11
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,0,1,1,4,4,5,5]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,0,1,1,4,4,5,5]
; AVX512DQBW-SLOW-NEXT: movabsq $-4165393823095705204, %rax # imm = 0xC6318C6318C6318C
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm7, %zmm11 {%k1}
-; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm16 = xmm7[0,0,1,1]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[0,0,0,1]
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[1,1,2,2]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,1,1]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm16, %zmm7
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm7, %zmm9 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,0,0,0,0,0,1,1,1,1,2,2,2,2,2,2]
+; AVX512DQBW-SLOW-NEXT: vpermd %zmm3, %zmm7, %zmm3
; AVX512DQBW-SLOW-NEXT: movabsq $595056260442243600, %rax # imm = 0x842108421084210
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm7, %zmm11 {%k1}
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm1, %ymm7
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm3, %zmm9 {%k1}
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm1, %ymm3
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm0, %ymm7
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,3,3]
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm0, %ymm16
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,3,3]
-; AVX512DQBW-SLOW-NEXT: vporq %ymm7, %ymm16, %ymm7
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm9, %ymm1, %ymm1
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm10, %ymm0, %ymm0
+; AVX512DQBW-SLOW-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm11, %ymm0, %ymm0
; AVX512DQBW-SLOW-NEXT: vpor %ymm1, %ymm0, %ymm0
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512DQBW-SLOW-NEXT: vpermd %ymm4, %ymm12, %ymm1
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm13, %ymm3, %ymm1 {%k3}
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm15, %ymm3, %ymm3
-; AVX512DQBW-SLOW-NEXT: vpshufb %ymm18, %ymm4, %ymm4
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512DQBW-SLOW-NEXT: vpermd %ymm5, %ymm12, %ymm1
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm14, %ymm4, %ymm1 {%k3}
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm17, %ymm4, %ymm3
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm18, %ymm5, %ymm4
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,3,3]
; AVX512DQBW-SLOW-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k2}
-; AVX512DQBW-SLOW-NEXT: vpermd %ymm5, %ymm14, %ymm0
-; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[0,2,1,1,4,6,5,5]
-; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,3,2]
-; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQBW-SLOW-NEXT: vpermd %ymm0, %ymm13, %ymm3
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,1,1,4,6,5,5]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,3,2]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512DQBW-SLOW-NEXT: movabsq $1190112520884487201, %rax # imm = 0x1084210842108421
; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, 64(%r9)
-; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm11, (%r9)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, (%r9)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm8, 256(%r9)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm6, 192(%r9)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm2, 128(%r9)
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 f6fb4837f19cff8..9c0dbdf03484b06 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
@@ -719,49 +719,91 @@ define void @store_i8_stride6_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
-; AVX512BW-LABEL: store_i8_stride6_vf16:
-; AVX512BW: # %bb.0:
-; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
-; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
-; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
-; AVX512BW-NEXT: vinserti32x4 $1, (%r9), %zmm2, %zmm2
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58,u,u]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58]
-; AVX512BW-NEXT: movl $1227105426, %ecx # imm = 0x49242492
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[0,2,0,2]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
-; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
-; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u]
-; AVX512BW-NEXT: movl $613566756, %ecx # imm = 0x24924924
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31,u,u]
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,3,11,u,u,u,u,4,12,u,u,u,u,5,13,u,u,u,u,22,30,u,u,u,u,23,31,u,u,u,u]
-; AVX512BW-NEXT: movw $18724, %cx # imm = 0x4924
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
-; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
-; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31]
-; AVX512BW-NEXT: movw $-28087, %cx # imm = 0x9249
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
-; AVX512BW-NEXT: vmovdqa %ymm0, 64(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512BW-NEXT: vzeroupper
-; AVX512BW-NEXT: retq
+; AVX512BW-SLOW-LABEL: store_i8_stride6_vf16:
+; AVX512BW-SLOW: # %bb.0:
+; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %xmm1
+; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm2
+; AVX512BW-SLOW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512BW-SLOW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512BW-SLOW-NEXT: vinserti32x4 $1, (%r9), %zmm2, %zmm2
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58,u,u]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58]
+; AVX512BW-SLOW-NEXT: movl $1227105426, %ecx # imm = 0x49242492
+; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[0,2,0,2]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
+; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u]
+; AVX512BW-SLOW-NEXT: movl $613566756, %ecx # imm = 0x24924924
+; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
+; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31,u,u]
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,3,11,u,u,u,u,4,12,u,u,u,u,5,13,u,u,u,u,22,30,u,u,u,u,23,31,u,u,u,u]
+; AVX512BW-SLOW-NEXT: movw $18724, %cx # imm = 0x4924
+; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
+; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
+; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31]
+; AVX512BW-SLOW-NEXT: movw $-28087, %cx # imm = 0x9249
+; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
+; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
+; AVX512BW-SLOW-NEXT: vmovdqa %ymm0, 64(%rax)
+; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm4, (%rax)
+; AVX512BW-SLOW-NEXT: vzeroupper
+; AVX512BW-SLOW-NEXT: retq
+;
+; AVX512BW-FAST-LABEL: store_i8_stride6_vf16:
+; AVX512BW-FAST: # %bb.0:
+; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512BW-FAST-NEXT: vmovdqa (%rdx), %xmm1
+; AVX512BW-FAST-NEXT: vmovdqa (%r8), %xmm2
+; AVX512BW-FAST-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
+; AVX512BW-FAST-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
+; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm3
+; AVX512BW-FAST-NEXT: vinserti32x4 $1, (%r9), %zmm2, %zmm2
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,2,0,2,8,10,9,11]
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm1, %zmm5
+; AVX512BW-FAST-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58,u,u]
+; AVX512BW-FAST-NEXT: vpermt2q %zmm1, %zmm4, %zmm3
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58]
+; AVX512BW-FAST-NEXT: movl $1227105426, %ecx # imm = 0x49242492
+; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm5, %zmm3 {%k1}
+; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,2,0,2,0,2,1,3]
+; AVX512BW-FAST-NEXT: vpermq %zmm2, %zmm4, %zmm4
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u]
+; AVX512BW-FAST-NEXT: movl $613566756, %ecx # imm = 0x24924924
+; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %zmm4, %zmm3 {%k1}
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31,u,u]
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,3,11,u,u,u,u,4,12,u,u,u,u,5,13,u,u,u,u,22,30,u,u,u,u,23,31,u,u,u,u]
+; AVX512BW-FAST-NEXT: movw $18724, %cx # imm = 0x4924
+; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
+; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
+; AVX512BW-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31]
+; AVX512BW-FAST-NEXT: movw $-28087, %cx # imm = 0x9249
+; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
+; AVX512BW-FAST-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
+; AVX512BW-FAST-NEXT: vmovdqa %ymm0, 64(%rax)
+; AVX512BW-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512BW-FAST-NEXT: vzeroupper
+; AVX512BW-FAST-NEXT: retq
%in.vec0 = load <16 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <16 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <16 x i8>, ptr %in.vecptr2, align 64
@@ -1788,99 +1830,94 @@ define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-FAST-LABEL: store_i8_stride6_vf32:
; AVX512F-FAST: # %bb.0:
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm11
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm12
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm13
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm8
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm9
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm1
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[16],ymm13[16],ymm12[17],ymm13[17],ymm12[18],ymm13[18],ymm12[19],ymm13[19],ymm12[20],ymm13[20],ymm12[21],ymm13[21],ymm12[22],ymm13[22],ymm12[23],ymm13[23]
-; AVX512F-FAST-NEXT: vprold $16, %ymm3, %ymm3
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm4
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm5
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[16],ymm11[16],ymm10[17],ymm11[17],ymm10[18],ymm11[18],ymm10[19],ymm11[19],ymm10[20],ymm11[20],ymm10[21],ymm11[21],ymm10[22],ymm11[22],ymm10[23],ymm11[23]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm14
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm2, %zmm14
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm3
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm6
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[16],ymm9[16],ymm8[17],ymm9[17],ymm8[18],ymm9[18],ymm8[19],ymm9[19],ymm8[20],ymm9[20],ymm8[21],ymm9[21],ymm8[22],ymm9[22],ymm8[23],ymm9[23]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,18,19,16,17,22,23,24,25,24,25,24,25,24,25]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm7
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm14 = [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]
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm11, %ymm15
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm10, %ymm14
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm14 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[4],ymm15[4],ymm14[5],ymm15[5],ymm14[6],ymm15[6],ymm14[7],ymm15[7],ymm14[16],ymm15[16],ymm14[17],ymm15[17],ymm14[18],ymm15[18],ymm14[19],ymm15[19],ymm14[20],ymm15[20],ymm14[21],ymm15[21],ymm14[22],ymm15[22],ymm14[23],ymm15[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15],ymm12[24],ymm13[24],ymm12[25],ymm13[25],ymm12[26],ymm13[26],ymm12[27],ymm13[27],ymm12[28],ymm13[28],ymm12[29],ymm13[29],ymm12[30],ymm13[30],ymm12[31],ymm13[31]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm15 = [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]
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm13, %ymm13
-; AVX512F-FAST-NEXT: vpshufb %ymm15, %ymm12, %ymm12
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[16],ymm13[16],ymm12[17],ymm13[17],ymm12[18],ymm13[18],ymm12[19],ymm13[19],ymm12[20],ymm13[20],ymm12[21],ymm13[21],ymm12[22],ymm13[22],ymm12[23],ymm13[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm10[8],ymm11[8],ymm10[9],ymm11[9],ymm10[10],ymm11[10],ymm10[11],ymm11[11],ymm10[12],ymm11[12],ymm10[13],ymm11[13],ymm10[14],ymm11[14],ymm10[15],ymm11[15],ymm10[24],ymm11[24],ymm10[25],ymm11[25],ymm10[26],ymm11[26],ymm10[27],ymm11[27],ymm10[28],ymm11[28],ymm10[29],ymm11[29],ymm10[30],ymm11[30],ymm10[31],ymm11[31]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm10
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512F-FAST-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-FAST-NEXT: vpternlogq $202, %zmm14, %zmm10, %zmm11
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm10 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm9, %ymm12
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm8, %ymm10
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[16],ymm12[16],ymm10[17],ymm12[17],ymm10[18],ymm12[18],ymm10[19],ymm12[19],ymm10[20],ymm12[20],ymm10[21],ymm12[21],ymm10[22],ymm12[22],ymm10[23],ymm12[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15],ymm8[24],ymm9[24],ymm8[25],ymm9[25],ymm8[26],ymm9[26],ymm8[27],ymm9[27],ymm8[28],ymm9[28],ymm8[29],ymm9[29],ymm8[30],ymm9[30],ymm8[31],ymm9[31]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,26,27,24,25,22,23,28,29,26,27,28,29,30,31]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm8
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm9 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm0, %xmm10
-; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm1, %xmm9
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm9, %zmm10, %zmm9
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm4, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm5, %xmm5
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm0
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm1 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm1
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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]
-; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[4,5,2,3,0,1,6,7,8,9,8,9,8,9,8,9]
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm1
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm2, %zmm1
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm4
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm5
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm1
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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]
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm3, %ymm7
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm2, %ymm6
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm7 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm5, %ymm8
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm4, %ymm7
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm7 = 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]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15],ymm2[24],ymm3[24],ymm2[25],ymm3[25],ymm2[26],ymm3[26],ymm2[27],ymm3[27],ymm2[28],ymm3[28],ymm2[29],ymm3[29],ymm2[30],ymm3[30],ymm2[31],ymm3[31]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512F-FAST-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-FAST-NEXT: vpternlogq $202, %zmm6, %zmm7, %zmm8
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm6 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm7
+; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm6
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm7 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,26,27,24,25,22,23,28,29,26,27,28,29,30,31]
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm9
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm7 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm9, %xmm8
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm10, %xmm7
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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]
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm7
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm11, %zmm12, %zmm11
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm13
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm12
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm12[8],xmm13[8],xmm12[9],xmm13[9],xmm12[10],xmm13[10],xmm12[11],xmm13[11],xmm12[12],xmm13[12],xmm12[13],xmm13[13],xmm12[14],xmm13[14],xmm12[15],xmm13[15]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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]
+; AVX512F-FAST-NEXT: vprold $16, %xmm13, %xmm13
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm13 = zmm12[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm11, %zmm13
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm11
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm14 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm11, %xmm15
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm12
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm12, %xmm14
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3],xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[4,5,2,3,0,1,6,7,8,9,8,9,8,9,8,9]
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm14, %zmm15, %zmm14
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm14 = zmm14[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm13, %zmm15, %zmm14
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[16],ymm5[16],ymm4[17],ymm5[17],ymm4[18],ymm5[18],ymm4[19],ymm5[19],ymm4[20],ymm5[20],ymm4[21],ymm5[21],ymm4[22],ymm5[22],ymm4[23],ymm5[23]
+; AVX512F-FAST-NEXT: vprold $16, %ymm4, %ymm4
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [0,0,0,1,10,10,10,11]
+; AVX512F-FAST-NEXT: vpermt2q %zmm4, %zmm9, %zmm5
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm2, %zmm9, %zmm3
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm15, %zmm3
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,18,19,16,17,22,23,24,25,24,25,24,25,24,25]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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]
+; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm9, %zmm1
+; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 64(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, (%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 128(%rax)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
@@ -4373,286 +4410,275 @@ define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512F-FAST-LABEL: store_i8_stride6_vf64:
; AVX512F-FAST: # %bb.0:
-; AVX512F-FAST-NEXT: subq $264, %rsp # imm = 0x108
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm11
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm13
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm6
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm7
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm8
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm9
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm4 = [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]
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm7, %ymm0
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm8, %ymm1
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm26
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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]
-; AVX512F-FAST-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm1
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm5, %ymm2
+; AVX512F-FAST-NEXT: subq $360, %rsp # imm = 0x168
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm4
+; AVX512F-FAST-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm3
+; AVX512F-FAST-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm3, %ymm2
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[16],ymm1[16],ymm2[17],ymm1[17],ymm2[18],ymm1[18],ymm2[19],ymm1[19],ymm2[20],ymm1[20],ymm2[21],ymm1[21],ymm2[22],ymm1[22],ymm2[23],ymm1[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm5[8],ymm3[8],ymm5[9],ymm3[9],ymm5[10],ymm3[10],ymm5[11],ymm3[11],ymm5[12],ymm3[12],ymm5[13],ymm3[13],ymm5[14],ymm3[14],ymm5[15],ymm3[15],ymm5[24],ymm3[24],ymm5[25],ymm3[25],ymm5[26],ymm3[26],ymm5[27],ymm3[27],ymm5[28],ymm3[28],ymm5[29],ymm3[29],ymm5[30],ymm3[30],ymm5[31],ymm3[31]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm5, %ymm27
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm28
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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]
-; AVX512F-FAST-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm2, %ymm2
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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]
-; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm9, %ymm3
-; AVX512F-FAST-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]
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
; AVX512F-FAST-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm9, %ymm5
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm9, %ymm29
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm5, %zmm3
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512F-FAST-NEXT: vmovdqa %ymm6, %ymm9
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm6, %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm5, %ymm4
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm5[8],ymm6[8],ymm5[9],ymm6[9],ymm5[10],ymm6[10],ymm5[11],ymm6[11],ymm5[12],ymm6[12],ymm5[13],ymm6[13],ymm5[14],ymm6[14],ymm5[15],ymm6[15],ymm5[24],ymm6[24],ymm5[25],ymm6[25],ymm5[26],ymm6[26],ymm5[27],ymm6[27],ymm5[28],ymm6[28],ymm5[29],ymm6[29],ymm5[30],ymm6[30],ymm5[31],ymm6[31]
-; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm7
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm11, %ymm0
-; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm13, %ymm6
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[4],ymm0[4],ymm6[5],ymm0[5],ymm6[6],ymm0[6],ymm6[7],ymm0[7],ymm6[16],ymm0[16],ymm6[17],ymm0[17],ymm6[18],ymm0[18],ymm6[19],ymm0[19],ymm6[20],ymm0[20],ymm6[21],ymm0[21],ymm6[22],ymm0[22],ymm6[23],ymm0[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm13[8],ymm11[8],ymm13[9],ymm11[9],ymm13[10],ymm11[10],ymm13[11],ymm11[11],ymm13[12],ymm11[12],ymm13[13],ymm11[13],ymm13[14],ymm11[14],ymm13[15],ymm11[15],ymm13[24],ymm11[24],ymm13[25],ymm11[25],ymm13[26],ymm11[26],ymm13[27],ymm11[27],ymm13[28],ymm11[28],ymm13[29],ymm11[29],ymm13[30],ymm11[30],ymm13[31],ymm11[31]
-; AVX512F-FAST-NEXT: vmovdqa %ymm11, %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm6
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm14
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm3, %ymm3
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm6
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm1 = [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]
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm6, %ymm3
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm7, %ymm4
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[1],ymm3[1],ymm4[2],ymm3[2],ymm4[3],ymm3[3],ymm4[4],ymm3[4],ymm4[5],ymm3[5],ymm4[6],ymm3[6],ymm4[7],ymm3[7],ymm4[16],ymm3[16],ymm4[17],ymm3[17],ymm4[18],ymm3[18],ymm4[19],ymm3[19],ymm4[20],ymm3[20],ymm4[21],ymm3[21],ymm4[22],ymm3[22],ymm4[23],ymm3[23]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm20
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm6, %ymm25
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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]
+; AVX512F-FAST-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm5
+; AVX512F-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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]
+; AVX512F-FAST-NEXT: # ymm7 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm5, %ymm4
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: # ymm9 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm5, %ymm5
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm14
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm14, %ymm6
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm8
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm8, %ymm0
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm8[8],ymm14[8],ymm8[9],ymm14[9],ymm8[10],ymm14[10],ymm8[11],ymm14[11],ymm8[12],ymm14[12],ymm8[13],ymm14[13],ymm8[14],ymm14[14],ymm8[15],ymm14[15],ymm8[24],ymm14[24],ymm8[25],ymm14[25],ymm8[26],ymm14[26],ymm8[27],ymm14[27],ymm8[28],ymm14[28],ymm8[29],ymm14[29],ymm8[30],ymm14[30],ymm8[31],ymm14[31]
+; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm6, %ymm2
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm6
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm6, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm11, %ymm1
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm6[8],ymm11[9],ymm6[9],ymm11[10],ymm6[10],ymm11[11],ymm6[11],ymm11[12],ymm6[12],ymm11[13],ymm6[13],ymm11[14],ymm6[14],ymm11[15],ymm6[15],ymm11[24],ymm6[24],ymm11[25],ymm6[25],ymm11[26],ymm6[26],ymm11[27],ymm6[27],ymm11[28],ymm6[28],ymm11[29],ymm6[29],ymm11[30],ymm6[30],ymm11[31],ymm6[31]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm14, %ymm0
-; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm14, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm2
+; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm16
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm4, %xmm1
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm2, %xmm2
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm2, %zmm24
-; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm10
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
-; AVX512F-FAST-NEXT: vpshufb %xmm6, %xmm10, %xmm0
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm10, %xmm11
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm11, %zmm25
-; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm0
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm15, %xmm11
-; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm0, %xmm12
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3],xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
-; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm12, %xmm1
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm11, %zmm1, %zmm23
-; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm1
-; AVX512F-FAST-NEXT: vpshufb %xmm6, %xmm1, %xmm6
-; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm2
-; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm6, %zmm2, %zmm22
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm13[0],ymm3[0],ymm13[1],ymm3[1],ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[4],ymm3[4],ymm13[5],ymm3[5],ymm13[6],ymm3[6],ymm13[7],ymm3[7],ymm13[16],ymm3[16],ymm13[17],ymm3[17],ymm13[18],ymm3[18],ymm13[19],ymm3[19],ymm13[20],ymm3[20],ymm13[21],ymm3[21],ymm13[22],ymm3[22],ymm13[23],ymm3[23]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm19
-; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm2
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm0[8],xmm15[8],xmm0[9],xmm15[9],xmm0[10],xmm15[10],xmm0[11],xmm15[11],xmm0[12],xmm15[12],xmm0[13],xmm15[13],xmm0[14],xmm15[14],xmm0[15],xmm15[15]
-; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm5[0],ymm9[0],ymm5[1],ymm9[1],ymm5[2],ymm9[2],ymm5[3],ymm9[3],ymm5[4],ymm9[4],ymm5[5],ymm9[5],ymm5[6],ymm9[6],ymm5[7],ymm9[7],ymm5[16],ymm9[16],ymm5[17],ymm9[17],ymm5[18],ymm9[18],ymm5[19],ymm9[19],ymm5[20],ymm9[20],ymm5[21],ymm9[21],ymm5[22],ymm9[22],ymm5[23],ymm9[23]
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm13
-; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm5
-; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm6
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm9, %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm30
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm27, %ymm3
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm28, %ymm9
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm9[0],ymm3[1],ymm9[1],ymm3[2],ymm9[2],ymm3[3],ymm9[3],ymm3[4],ymm9[4],ymm3[5],ymm9[5],ymm3[6],ymm9[6],ymm3[7],ymm9[7],ymm3[16],ymm9[16],ymm3[17],ymm9[17],ymm3[18],ymm9[18],ymm3[19],ymm9[19],ymm3[20],ymm9[20],ymm3[21],ymm9[21],ymm3[22],ymm9[22],ymm3[23],ymm9[23]
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm28
-; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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]
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm11, %xmm12
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm3
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm31
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm26, %ymm3
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-FAST-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm15, %ymm9
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm3, %ymm3
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm3, %ymm20
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm11
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm10, %xmm1
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm9
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm9, %xmm1
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm7
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm7, %xmm2
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3],xmm7[4],xmm9[4],xmm7[5],xmm9[5],xmm7[6],xmm9[6],xmm7[7],xmm9[7]
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
+; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm2, %xmm2
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm26
+; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm4
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm12 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm4, %xmm2
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm4, %xmm3
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm4, %xmm17
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm27
+; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm5, %xmm10
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm10[8],xmm0[9],xmm10[9],xmm0[10],xmm10[10],xmm0[11],xmm10[11],xmm0[12],xmm10[12],xmm0[13],xmm10[13],xmm0[14],xmm10[14],xmm0[15],xmm10[15]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
+; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm10, %xmm10
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm10, %zmm28
+; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm13
+; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm13, %xmm0
+; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm13, %xmm1
+; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm29
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm12
+; AVX512F-FAST-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]
; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm14, %ymm8
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm29, %ymm3
-; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm3, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm29
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm3
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-FAST-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm3, %ymm1
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm10
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm15
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm7
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm18
-; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [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]
-; AVX512F-FAST-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm10, %ymm7
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm16
-; AVX512F-FAST-NEXT: vpshufb %ymm4, %ymm3, %ymm4
-; AVX512F-FAST-NEXT: vmovdqa64 %ymm4, %ymm17
-; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm3, %ymm14
-; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm6, %xmm10
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm10[0],xmm4[0],xmm10[1],xmm4[1],xmm10[2],xmm4[2],xmm10[3],xmm4[3],xmm10[4],xmm4[4],xmm10[5],xmm4[5],xmm10[6],xmm4[6],xmm10[7],xmm4[7]
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512F-FAST-NEXT: vprold $16, %xmm5, %xmm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm26
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm2, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm0, %xmm3
-; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm27
-; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm2
-; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm3
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm10
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm3, %xmm5
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm7
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm3, %xmm6
-; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm4
-; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm3
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm13[0,0,0,1]
-; AVX512F-FAST-NEXT: vprold $16, %ymm19, %ymm2
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm1[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm21 = ymm0[2,2,2,3]
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm22 = ymm0[2,2,2,3]
+; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm4
+; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm10 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm4, %xmm0
+; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm3, %xmm15
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3],xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; AVX512F-FAST-NEXT: vprold $16, %xmm15, %xmm15
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm23
+; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm0
+; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm15
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm0, %xmm1
+; AVX512F-FAST-NEXT: vpshufb %xmm10, %xmm15, %xmm10
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm1[0],xmm10[1],xmm1[1],xmm10[2],xmm1[2],xmm10[3],xmm1[3],xmm10[4],xmm1[4],xmm10[5],xmm1[5],xmm10[6],xmm1[6],xmm10[7],xmm1[7]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3],xmm15[4],xmm0[4],xmm15[5],xmm0[5],xmm15[6],xmm0[6],xmm15[7],xmm0[7]
+; AVX512F-FAST-NEXT: vprold $16, %xmm10, %xmm10
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm10, %zmm24
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm11[0],ymm6[0],ymm11[1],ymm6[1],ymm11[2],ymm6[2],ymm11[3],ymm6[3],ymm11[4],ymm6[4],ymm11[5],ymm6[5],ymm11[6],ymm6[6],ymm11[7],ymm6[7],ymm11[16],ymm6[16],ymm11[17],ymm6[17],ymm11[18],ymm6[18],ymm11[19],ymm6[19],ymm11[20],ymm6[20],ymm11[21],ymm6[21],ymm11[22],ymm6[22],ymm11[23],ymm6[23]
+; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm15[8],xmm0[8],xmm15[9],xmm0[9],xmm15[10],xmm0[10],xmm15[11],xmm0[11],xmm15[12],xmm0[12],xmm15[13],xmm0[13],xmm15[14],xmm0[14],xmm15[15],xmm0[15]
+; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm8[0],ymm14[0],ymm8[1],ymm14[1],ymm8[2],ymm14[2],ymm8[3],ymm14[3],ymm8[4],ymm14[4],ymm8[5],ymm14[5],ymm8[6],ymm14[6],ymm8[7],ymm14[7],ymm8[16],ymm14[16],ymm8[17],ymm14[17],ymm8[18],ymm14[18],ymm8[19],ymm14[19],ymm8[20],ymm14[20],ymm8[21],ymm14[21],ymm8[22],ymm14[22],ymm8[23],ymm14[23]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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]
+; AVX512F-FAST-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm0
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm11 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
+; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm2, %xmm6
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm31 = [0,0,0,1,10,10,10,11]
+; AVX512F-FAST-NEXT: vpermt2q %zmm0, %zmm31, %zmm6
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm10, %xmm2
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm31, %zmm2
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm30, %zmm6
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm16, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
+; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm13, %xmm2
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm8, %xmm16
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm31, %zmm2
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm20, %ymm1
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm25, %ymm8
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
+; AVX512F-FAST-NEXT: vprold $16, %ymm1, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %xmm0, %xmm3, %xmm3
+; AVX512F-FAST-NEXT: vpermt2q %zmm1, %zmm31, %zmm3
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
+; AVX512F-FAST-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]
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm13
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm19, %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm13, %ymm8
+; AVX512F-FAST-NEXT: vmovdqa64 %ymm18, %ymm0
+; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm13, %ymm1
+; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm4, %ymm4
+; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm15
+; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
+; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm7
+; AVX512F-FAST-NEXT: vpshufb %xmm11, %xmm5, %xmm5
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
+; AVX512F-FAST-NEXT: vpermt2q %zmm4, %zmm31, %zmm5
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm7, %xmm4
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm30, %zmm5
+; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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]
+; AVX512F-FAST-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm13, %ymm11
+; AVX512F-FAST-NEXT: vpshufb %ymm3, %ymm12, %ymm9
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
+; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm15, %xmm12
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm13
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm14
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm15
+; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm7, %xmm0
+; AVX512F-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm10
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm2[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm12, %zmm2
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm9, %zmm2
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm2, %ymm0
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm0, %ymm1, %ymm13
-; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm10
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm2
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm10 = zmm10[0,1,2,3],zmm2[4,5,6,7]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm2
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm10, %zmm8, %zmm2
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm30[0,0,0,1]
-; AVX512F-FAST-NEXT: vprold $16, %ymm28, %ymm11
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm4[0,0,0,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm13[0,0,0,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm12[0,0,0,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm14[0,0,0,1]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm15[0,0,0,1]
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm17, %xmm14
+; AVX512F-FAST-NEXT: vmovdqa64 %xmm16, %xmm15
+; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm14, %xmm14
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm6, %ymm15, %ymm13
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm6, %ymm6
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm31[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm20[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm15[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm21[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm29[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm18[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm16[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm17[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm11
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm10, %zmm9, %zmm11
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm10
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm10, %ymm1, %ymm13
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm11, %ymm0, %ymm5
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm10
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm10[4,5,6,7]
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm28, %zmm15, %zmm10
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm8, %zmm10
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm15, %ymm0
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm5, %ymm5
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512F-FAST-NEXT: vpermt2q %zmm10, %zmm31, %zmm14
+; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm10, %ymm9
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm5
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm5[4,5,6,7]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm14
+; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm14, 256(%rax)
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm6, %ymm10, %ymm11
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm13[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm2
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 64(%rax)
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm30, %zmm2
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm8 = mem[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm8
-; AVX512F-FAST-NEXT: vpermq $234, (%rsp), %zmm5 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm11 = mem[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm11
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm8, %ymm5
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm0, %ymm29
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm5
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm0, %ymm31
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm0, %zmm30, %zmm5
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm2, %ymm0
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm0, %ymm15, %ymm21
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm5, %ymm0
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm0, %ymm15, %ymm8
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm8, %ymm0, %ymm30
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm0, %zmm5
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm30[0,1,2,3],zmm5[4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm8 = mem[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm2, %ymm0, %ymm22
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm21, %zmm0, %zmm2
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm22[0,1,2,3],zmm2[4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [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]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm9, %zmm8
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm11, %ymm0, %ymm14
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm31, %zmm0, %zmm5
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm14[0,1,2,3],zmm5[4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11 # 64-byte Folded Reload
-; AVX512F-FAST-NEXT: # zmm11 = mem[2,2,2,3,6,6,6,7]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm9, %zmm11
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm2, %zmm9, %zmm6
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm0, %ymm1
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm2
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm2[4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
+; AVX512F-FAST-NEXT: # zmm2 = mem[2,2,2,3,6,6,6,7]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm1, %zmm9, %zmm2
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm23[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm5 = zmm26[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm9 = zmm24[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm12, %zmm9
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm5 = zmm27[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm13 = zmm23[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm12, %zmm13
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm5
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm0, %ymm7
-; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm13, %ymm5
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm0, %ymm4
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm9, %ymm1, %ymm6
-; AVX512F-FAST-NEXT: vpternlogq $184, %ymm13, %ymm1, %ymm3
-; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm6[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm25[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-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]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm1
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm8, %zmm5
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm24[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm9 = zmm28[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vpternlogq $226, %zmm1, %zmm8, %zmm9
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm5, %ymm1
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm1, %ymm0, %ymm3
+; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm1
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm1, %ymm0, %ymm4
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm10, %ymm7
+; AVX512F-FAST-NEXT: vpternlogq $184, %ymm9, %ymm10, %ymm12
+; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm7[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm27[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [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]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
-; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm3 = zmm22[0,0,0,1,4,4,4,5]
-; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm3
-; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm12[0,1,2,3],zmm0[4,5,6,7]
+; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm4 = zmm29[0,0,0,1,4,4,4,5]
+; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm4
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, 192(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm11, 128(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 320(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm10, 256(%rax)
-; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 64(%rax)
-; AVX512F-FAST-NEXT: addq $264, %rsp # imm = 0x108
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 128(%rax)
+; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 320(%rax)
+; AVX512F-FAST-NEXT: addq $360, %rsp # imm = 0x168
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-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 0ec4e796fcfacaa..97be882ffd8e43f 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
@@ -3316,159 +3316,160 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
-; AVX512F-SLOW-LABEL: store_i8_stride7_vf32:
-; AVX512F-SLOW: # %bb.0:
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512F-SLOW-NEXT: vmovdqa64 (%rdx), %ymm18
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm5
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm6
-; AVX512F-SLOW-NEXT: vmovdqa (%r10), %ymm4
-; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm8
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm8[u],zero,xmm8[7],zero,xmm8[5,u,u,u],zero,xmm8[8],zero,xmm8[6,u,u,u],zero
-; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm9[u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u,u,9]
-; AVX512F-SLOW-NEXT: vpor %xmm7, %xmm10, %xmm7
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm7, %zmm10, %zmm7
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm13
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm13[u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6,u,u]
-; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm14
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm11 = xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero,xmm14[u,u]
-; AVX512F-SLOW-NEXT: vpor %xmm10, %xmm11, %xmm10
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm10, %zmm11, %zmm10
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm15 = zmm10[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm15
-; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm10
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = zero,xmm10[4,u,u,u],zero,xmm10[7],zero,xmm10[5,u,u,u],zero,xmm10[8],zero,xmm10[6]
-; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm11
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm12 = xmm11[4],zero,xmm11[u,u,u,7],zero,xmm11[5],zero,xmm11[u,u,u,8],zero,xmm11[6],zero
-; AVX512F-SLOW-NEXT: vpor %xmm7, %xmm12, %xmm7
-; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm7, %zmm12, %zmm7
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm16 = zmm7[0,1,0,1,4,5,4,5]
-; AVX512F-SLOW-NEXT: vmovdqa (%r10), %xmm12
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm12[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm12[1,1,0,0,4,5,6,7]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm0[0,0,1,0,4,4,5,4]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm16, %zmm7
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm15, %zmm7
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14,u,u],zero,zero,zero,zero,ymm1[15,u,u],zero,zero,zero,zero,ymm1[16,u,u],zero,zero,zero,zero,ymm1[17,u,u],zero,zero,zero,zero,ymm1[18]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm15 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm15, %ymm0
-; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm14[8],xmm13[9],xmm14[9],xmm13[10],xmm14[10],xmm13[11],xmm14[11],xmm13[12],xmm14[12],xmm13[13],xmm14[13],xmm13[14],xmm14[14],xmm13[15],xmm14[15]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm13, %zmm0
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u,u,u]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm18, %ymm2
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm14 = ymm2[u,u,u,u,14],zero,ymm2[u,u,u,u,u,15],zero,ymm2[u,u,u,u,u,16],zero,ymm2[u,u,u,u,u,17],zero,ymm2[u,u,u,u,u]
-; AVX512F-SLOW-NEXT: vpor %ymm13, %ymm14, %ymm13
-; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm8, %zmm9
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm9
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm5[u,u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[13,u,u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u]
-; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm8, %ymm0
-; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm8 = xmm12[0,1,2,3,4,5,5,6]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[2,2,3,3]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,0,1]
-; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
-; AVX512F-SLOW-NEXT: vpandn %ymm8, %ymm10, %ymm8
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = zero,ymm4[13,u,u,u,u],zero,zero,ymm4[14,u,u,u,u],zero,zero,ymm4[15,u,u,u,u],zero,zero,ymm4[16,u,u,u,u],zero,zero,ymm4[17,u,u]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm9 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
-; AVX512F-SLOW-NEXT: # ymm9 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpand %ymm0, %ymm9, %ymm0
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm2[18,19,20,21],zero,ymm2[19],zero,ymm2[25,26,27,22],zero,ymm2[20],zero
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm10, %zmm0
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20]
-; AVX512F-SLOW-NEXT: vmovdqa64 %ymm16, %ymm14
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero,zero,zero,ymm14[27],zero,ymm14[25]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vporq %zmm10, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm14[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,1,1,4,4,5,5]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512F-SLOW-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512F-SLOW-NEXT: vpand %ymm11, %ymm10, %ymm10
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = 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]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm13 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm12 = zmm12[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vporq %zmm12, %zmm10, %zmm10
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm10
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm0
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm12 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm5[18],zero,ymm5[20,21,20,21],zero,ymm5[19],zero,ymm5[19,20,21,22],zero
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm13 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm5[23],zero,ymm5[23,24,25,26],zero,ymm5[24],zero,ymm5[30,31]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm12 = zmm12[2,3,2,3,6,7,6,7]
-; AVX512F-SLOW-NEXT: vporq %zmm0, %zmm12, %zmm0
-; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,1,1,3,4,5,5,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,3,2]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm13 = 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]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
-; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm12
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm12
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm9, %ymm0, %ymm5
-; AVX512F-SLOW-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]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm0
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-SLOW-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
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm11, %ymm3, %ymm2
-; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm14[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
-; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm3
-; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
-; AVX512F-SLOW-NEXT: vmovdqa %ymm3, 192(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm7, (%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm12, 128(%rax)
-; AVX512F-SLOW-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512F-SLOW-NEXT: vzeroupper
-; AVX512F-SLOW-NEXT: retq
+; AVX512F-ONLY-SLOW-LABEL: store_i8_stride7_vf32:
+; AVX512F-ONLY-SLOW: # %bb.0:
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm5
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm6
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r10), %ymm4
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,5,4,u,5,u,4,u,20,21,u,23,u,21,u,23>
+; AVX512F-ONLY-SLOW-NEXT: vpermi2d %zmm7, %zmm8, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm5[18],zero,ymm5[20,21,20,21],zero,ymm5[19],zero,ymm5[19,20,21,22],zero
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm5[23],zero,ymm5[23,24,25,26],zero,ymm5[24],zero,ymm5[30,31]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vporq %zmm7, %zmm8, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm16 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
+; AVX512F-ONLY-SLOW-NEXT: # ymm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpandq %ymm16, %ymm8, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm11[18,19,20,21],zero,ymm11[19],zero,ymm11[25,26,27,22],zero,ymm11[20],zero
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vporq %zmm9, %zmm8, %zmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm17 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
+; AVX512F-ONLY-SLOW-NEXT: # ymm17 = mem[0,1,2,3,0,1,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpandq %ymm17, %ymm8, %ymm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm11[23],zero,ymm11[21,22,23,26],zero,ymm11[24],zero,ymm11[28,29,26,27]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm11, %ymm20
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vporq %zmm10, %zmm8, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm11
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm11[u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6,u,u,u],zero
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm12
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm9 = xmm12[u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u,u,9]
+; AVX512F-ONLY-SLOW-NEXT: vpor %xmm7, %xmm9, %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm7, %zmm9, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm7[0,1,0,1,4,5,4,5]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[u,u,u],zero,xmm7[7],zero,xmm7[5,u,u,u],zero,xmm7[8],zero,xmm7[6,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm9[u,u,u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u]
+; AVX512F-ONLY-SLOW-NEXT: vpor %xmm13, %xmm14, %xmm13
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm13, %zmm14, %zmm13
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm18 = zmm13[0,1,0,1,4,5,4,5]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm18
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm13
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = zero,xmm13[4,u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm14
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm15 = xmm14[4],zero,xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero
+; AVX512F-ONLY-SLOW-NEXT: vpor %xmm10, %xmm15, %xmm10
+; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm19 = zmm10[0,1,0,1,4,5,4,5]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r10), %xmm15
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm15[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm15[1,1,0,0,4,5,6,7]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm10
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14,u,u],zero,zero,zero,zero,ymm1[15,u,u],zero,zero,zero,zero,ymm1[16,u,u],zero,zero,zero,zero,ymm1[17,u,u],zero,zero,zero,zero,ymm1[18]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm20, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,14],zero,ymm2[u,u,u,u,u,15],zero,ymm2[u,u,u,u,u,16],zero,ymm2[u,u,u,u,u,17],zero,ymm2[u,u,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm1, %ymm7, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm5[u,u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[13,u,u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm15[0,1,2,3,4,5,5,6]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
+; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm7, %ymm9, %ymm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm4[13,u,u,u,u],zero,zero,ymm4[14,u,u,u,u],zero,zero,ymm4[15,u,u,u,u],zero,zero,ymm4[16,u,u,u,u],zero,zero,ymm4[17,u,u]
+; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, %ymm16, %ymm0, %ymm1
+; AVX512F-ONLY-SLOW-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]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512F-ONLY-SLOW-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
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, %ymm17, %ymm1, %ymm2
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm19, %ymm1
+; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
+; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm3, 192(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm10, (%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm8, 128(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512F-ONLY-SLOW-NEXT: vzeroupper
+; AVX512F-ONLY-SLOW-NEXT: retq
;
; AVX512F-FAST-LABEL: store_i8_stride7_vf32:
; AVX512F-FAST: # %bb.0:
@@ -3619,6 +3620,161 @@ define void @store_i8_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
+; AVX512DQ-SLOW-LABEL: store_i8_stride7_vf32:
+; AVX512DQ-SLOW: # %bb.0:
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm11
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm5
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm6
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r10), %ymm4
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,u,u,u,u,26,27,24,25]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm4[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = <u,5,4,u,5,u,4,u,20,21,u,23,u,21,u,23>
+; AVX512DQ-SLOW-NEXT: vpermi2d %zmm7, %zmm8, %zmm9
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm6[20],zero,ymm6[18],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[25],zero,ymm6[23],zero,zero,zero,zero,ymm6[26],zero,ymm6[24],zero,zero
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm7
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm7[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,ymm5[18],zero,ymm5[20,21,20,21],zero,ymm5[19],zero,ymm5[19,20,21,22],zero
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,ymm5[23],zero,ymm5[23,24,25,26],zero,ymm5[24],zero,ymm5[30,31]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vporq %zmm7, %zmm8, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm7
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,25,u,23,u,u,u,u,26,u,24,u,u,u,u,27,u]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm16 = [18374967954648269055,71777218572844800,18374967954648269055,71777218572844800]
+; AVX512DQ-SLOW-NEXT: # ymm16 = mem[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpandq %ymm16, %ymm8, %ymm8
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm11[18,19,20,21],zero,ymm11[19],zero,ymm11[25,26,27,22],zero,ymm11[20],zero
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm3[18],zero,zero,zero,zero,ymm3[21],zero,ymm3[19],zero,zero,zero,zero,ymm3[22],zero,ymm3[20]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero,zero,zero,ymm2[27],zero,ymm2[25]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vporq %zmm9, %zmm8, %zmm9
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm8 = ymm2[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,0,1,1,4,4,5,5]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm17 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
+; AVX512DQ-SLOW-NEXT: # ymm17 = mem[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vpandq %ymm17, %ymm8, %ymm8
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm11[23],zero,ymm11[21,22,23,26],zero,ymm11[24],zero,ymm11[28,29,26,27]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm11, %ymm20
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20],zero,zero
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm3[25],zero,ymm3[23],zero,zero,zero,zero,ymm3[26],zero,ymm3[24],zero,zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm10, %zmm10
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,3,2,3,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vporq %zmm10, %zmm8, %zmm8
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm9, %zmm8
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm8
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm11
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm11[u],zero,xmm11[7],zero,xmm11[5,u,u,u],zero,xmm11[8],zero,xmm11[6,u,u,u],zero
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm12
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm9 = xmm12[u,7],zero,xmm12[5],zero,xmm12[u,u,u,8],zero,xmm12[6],zero,xmm12[u,u,u,9]
+; AVX512DQ-SLOW-NEXT: vpor %xmm7, %xmm9, %xmm7
+; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
+; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm7, %zmm9, %zmm7
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm7[0,1,0,1,4,5,4,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm7
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[u,u,u],zero,xmm7[7],zero,xmm7[5,u,u,u],zero,xmm7[8],zero,xmm7[6,u,u]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm9
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm9[u,u,u,7],zero,xmm9[5],zero,xmm9[u,u,u,8],zero,xmm9[6],zero,xmm9[u,u]
+; AVX512DQ-SLOW-NEXT: vpor %xmm13, %xmm14, %xmm13
+; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
+; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm13, %zmm14, %zmm13
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm18 = zmm13[0,1,0,1,4,5,4,5]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm10, %zmm18
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm13
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = zero,xmm13[4,u,u,u],zero,xmm13[7],zero,xmm13[5,u,u,u],zero,xmm13[8],zero,xmm13[6]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm14
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm15 = xmm14[4],zero,xmm14[u,u,u,7],zero,xmm14[5],zero,xmm14[u,u,u,8],zero,xmm14[6],zero
+; AVX512DQ-SLOW-NEXT: vpor %xmm10, %xmm15, %xmm10
+; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm19 = zmm10[0,1,0,1,4,5,4,5]
+; AVX512DQ-SLOW-NEXT: vmovdqa (%r10), %xmm15
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm15[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm15[1,1,0,0,4,5,6,7]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,2,0]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm10 = zmm0[0,0,1,0,4,4,5,4]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm19, %zmm10
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm18, %zmm10
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14,u,u],zero,zero,zero,zero,ymm1[15,u,u],zero,zero,zero,zero,ymm1[16,u,u],zero,zero,zero,zero,ymm1[17,u,u],zero,zero,zero,zero,ymm1[18]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm2[0,1,14],zero,ymm2[u,u,0,1,14,15],zero,ymm2[u,u,13,2,3,16],zero,ymm2[u,u,28,29,16,17],zero,ymm2[u,u,19,28,29,18],zero
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm18
+; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u],zero,ymm3[14,u,u,u,u,u],zero,ymm3[15,u,u,u,u,u],zero,ymm3[16,u,u,u,u,u],zero,ymm3[17,u,u,u,u,u]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm20, %ymm2
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm2[u,u,u,u,14],zero,ymm2[u,u,u,u,u,15],zero,ymm2[u,u,u,u,u,16],zero,ymm2[u,u,u,u,u,17],zero,ymm2[u,u,u,u,u]
+; AVX512DQ-SLOW-NEXT: vpor %ymm1, %ymm7, %ymm1
+; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm1
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm5[u,u,u,u,u,14],zero,ymm5[u,u,u,u,u,15],zero,ymm5[u,u,u,u,u,16],zero,ymm5[u,u,u,u,u,17],zero,ymm5[u,u,u]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[13,u,u,u,u,u],zero,ymm6[14,u,u,u,u,u],zero,ymm6[15,u,u,u,u,u],zero,ymm6[16,u,u,u,u,u],zero,ymm6[17,u,u,u]
+; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm7, %ymm0
+; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm7, %zmm0
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm7 = xmm15[0,1,2,3,4,5,5,6]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[2,2,3,3]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
+; AVX512DQ-SLOW-NEXT: vpandn %ymm7, %ymm9, %ymm7
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = zero,ymm4[13,u,u,u,u],zero,zero,ymm4[14,u,u,u,u],zero,zero,ymm4[15,u,u,u,u],zero,zero,ymm4[16,u,u,u,u],zero,zero,ymm4[17,u,u]
+; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm7, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm7
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm5[27],zero,zero,zero,zero,ymm5[30],zero,ymm5[28],zero,zero,zero,zero,ymm5[31],zero,ymm5[29]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, %ymm16, %ymm0, %ymm1
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm0
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,27,u,u,u,u,30,u,28,u,u,u,u,31,u,29,u]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512DQ-SLOW-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-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpternlogq $248, %ymm17, %ymm1, %ymm2
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm19, %ymm1
+; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm18, %ymm3
+; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
+; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
+; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm3
+; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm3
+; AVX512DQ-SLOW-NEXT: vmovdqa %ymm3, 192(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm10, (%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm8, 128(%rax)
+; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
+; AVX512DQ-SLOW-NEXT: vzeroupper
+; AVX512DQ-SLOW-NEXT: retq
+;
; AVX512BW-SLOW-LABEL: store_i8_stride7_vf32:
; AVX512BW-SLOW: # %bb.0:
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
@@ -7222,446 +7378,443 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
-; AVX512F-ONLY-SLOW-LABEL: store_i8_stride7_vf64:
-; AVX512F-ONLY-SLOW: # %bb.0:
-; AVX512F-ONLY-SLOW-NEXT: subq $1288, %rsp # imm = 0x508
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm1, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm2, %ymm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm2, %ymm6, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm2
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512F-ONLY-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm22
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %ymm12
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm12[27],zero,zero,zero,zero,ymm12[30],zero,ymm12[28],zero,zero,zero,zero,ymm12[31],zero,ymm12[29]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-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
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29],zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm3, %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm9, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm9, %ymm25
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rax), %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm7, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[18],zero,zero,zero,zero,ymm7[21],zero,ymm7[19],zero,zero,zero,zero,ymm7[22],zero,ymm7[20]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm6[23],zero,ymm6[21,22,23,26],zero,ymm6[24],zero,ymm6[28,29,26,27]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm6[18,19,20,21],zero,ymm6[19],zero,ymm6[25,26,27,22],zero,ymm6[20],zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm6, %ymm27
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22],zero,ymm5[20],zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm1, %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm3, %xmm19
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm1, %xmm31
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm2, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm3, %xmm24
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm2, %xmm30
-; AVX512F-ONLY-SLOW-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm1, %xmm6, %xmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm0, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpor %xmm4, %xmm5, %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm11
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm11, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm2, %xmm4, %xmm9
-; AVX512F-ONLY-SLOW-NEXT: vpor %xmm5, %xmm9, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm21, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm5, %ymm10, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm10, %ymm17
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm16, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm14, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero,zero,zero,ymm8[18]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm8, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm25, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm14[0,1,14],zero,ymm14[12,13,0,1,14,15],zero,ymm14[3,12,13,2,3,16],zero,ymm14[30,31,28,29,16,17],zero,ymm14[31,18,19,28,29,18],zero
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm23, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm5, %ymm12, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm18, %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm8, %ymm15, %ymm9
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm15, %ymm25
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm1, %xmm5, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm5, %xmm15
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm9, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vporq %xmm1, %xmm3, %xmm21
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm1, %xmm12, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm3, %xmm10, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpor %xmm1, %xmm3, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm5, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm2, %xmm8, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpor %xmm1, %xmm3, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm5[8],xmm8[9],xmm5[9],xmm8[10],xmm5[10],xmm8[11],xmm5[11],xmm8[12],xmm5[12],xmm8[13],xmm5[13],xmm8[14],xmm5[14],xmm8[15],xmm5[15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm2, %xmm16
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rax), %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,6]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
-; AVX512F-ONLY-SLOW-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa (%rax), %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = zero,ymm2[13],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
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm22
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm1, %ymm23
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm28, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm3, %ymm29
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[27],zero,zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm14[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm14, %ymm24
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm7, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm7, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm26
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,1,1,4,4,5,5]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm27
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm30, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm31, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm5, %xmm19
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm30
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512F-ONLY-SLOW-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]
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm15[8],xmm9[8],xmm15[9],xmm9[9],xmm15[10],xmm9[10],xmm15[11],xmm9[11],xmm15[12],xmm9[12],xmm15[13],xmm9[13],xmm15[14],xmm9[14],xmm15[15],xmm9[15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %xmm15, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm6 = <2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm6, %xmm3, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm6, %xmm0, %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm31
-; AVX512F-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm11[8],xmm4[9],xmm11[9],xmm4[10],xmm11[10],xmm4[11],xmm11[11],xmm4[12],xmm11[12],xmm4[13],xmm11[13],xmm4[14],xmm11[14],xmm4[15],xmm11[15]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm17, %ymm11
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm11, %ymm6
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22],zero,ymm11[20]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm15[23],zero,ymm15[21,22,23,26],zero,ymm15[24],zero,ymm15[28,29,26,27]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm15[18,19,20,21],zero,ymm15[19],zero,ymm15[25,26,27,22],zero,ymm15[20],zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[21],zero,ymm14[19],zero,zero,zero,zero,ymm14[22],zero,ymm14[20],zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm25, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm29, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm0, %ymm14, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm2, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm0, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm14, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm15, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm13, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm2, %ymm29
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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]
-; AVX512F-ONLY-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm13, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %ymm4, %ymm15, %ymm2
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = zero,ymm2[13],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
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm2[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,2]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255]
-; AVX512F-ONLY-SLOW-NEXT: vpandnq %ymm15, %ymm25, %ymm15
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm4, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5>
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm7, %xmm15
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm9, %xmm5
-; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm21, %zmm5, %zmm25
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm1[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm23[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm12[0],xmm10[1],xmm12[1],xmm10[2],xmm12[2],xmm10[3],xmm12[3],xmm10[4],xmm12[4],xmm10[5],xmm12[5],xmm10[6],xmm12[6],xmm10[7],xmm12[7]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9>
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %xmm19, %xmm2
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm12
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm8[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,3,3,6,6,7,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm3[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm11[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm18[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm16[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm17[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %ymm24, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,0,1,1,4,4,5,5]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm1, %xmm8
-; AVX512F-ONLY-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vporq %zmm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512F-ONLY-SLOW-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpandq %ymm19, %ymm6, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vporq %zmm2, %zmm0, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm3 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [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]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm3, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vporq %zmm1, %zmm2, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm27[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm3, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm6 = zmm30[0,1,0,1,4,5,4,5]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm31[0,1,0,1,4,5,4,5]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm3, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, %ymm19, %ymm21, %ymm14
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm7[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, %ymm19, %ymm10, %ymm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm14, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm9[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm20, %ymm7
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $184, %zmm6, %zmm9, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm3
-; AVX512F-ONLY-SLOW-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm19 # 16-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm28[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm5[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]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # xmm10 = xmm10[0],mem[0],xmm10[1],mem[1],xmm10[2],mem[2],xmm10[3],mem[3],xmm10[4],mem[4],xmm10[5],mem[5],xmm10[6],mem[6],xmm10[7],mem[7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa 32(%rax), %xmm11
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm21 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm21 = zmm10[0,1,0,1],mem[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm10[1,1,0,0,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,1,2,0]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm10, %xmm14
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm12
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm11[1,1,0,0,4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[0,1,2,0]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm15 = mem[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm29[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm17 = mem[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm18 = mem[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa %ymm5, %ymm10
-; AVX512F-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[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]
-; AVX512F-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm11, %xmm13
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,1,0]
-; AVX512F-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,5,5,6]
-; AVX512F-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, (%rsp), %ymm23 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm23 = mem[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # ymm26 = mem[2,3,2,3]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm22
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm22
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm14[0,0,1,0]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm6
-; AVX512F-ONLY-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512F-ONLY-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
-; AVX512F-ONLY-SLOW-NEXT: vporq %ymm15, %ymm16, %ymm0
-; AVX512F-ONLY-SLOW-NEXT: vporq %ymm17, %ymm18, %ymm8
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm0[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm10[2,2,3,2]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm5
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm11[0,1,0,1]
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm13, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512F-ONLY-SLOW-NEXT: vporq %ymm23, %ymm26, %ymm1
-; AVX512F-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512F-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm25[0,1,0,1,4,5,4,5]
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm19[0,1,0,1,4,5,4,5]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
-; AVX512F-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm12[0,0,1,0,4,4,5,4]
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
-; AVX512F-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm4, 320(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512F-ONLY-SLOW-NEXT: vmovdqa64 %zmm22, 64(%rax)
-; AVX512F-ONLY-SLOW-NEXT: addq $1288, %rsp # imm = 0x508
-; AVX512F-ONLY-SLOW-NEXT: vzeroupper
-; AVX512F-ONLY-SLOW-NEXT: retq
+; AVX512F-SLOW-LABEL: store_i8_stride7_vf64:
+; AVX512F-SLOW: # %bb.0:
+; AVX512F-SLOW-NEXT: subq $1416, %rsp # imm = 0x588
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
+; AVX512F-SLOW-NEXT: vmovdqa %ymm1, %ymm12
+; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm2
+; AVX512F-SLOW-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
+; AVX512F-SLOW-NEXT: vmovdqa %ymm2, %ymm9
+; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm27
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm8
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm8, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm28
+; AVX512F-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero,zero,zero
+; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm2
+; AVX512F-SLOW-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]
+; AVX512F-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm3, %ymm17
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512F-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm23
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %ymm10
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm11
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29]
+; AVX512F-SLOW-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm10[25],zero,ymm10[23],zero,zero,zero,zero,ymm10[26],zero,ymm10[24],zero,zero
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm5
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm6
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[30],zero,ymm6[28],zero,zero,zero,zero,ymm6[31],zero,ymm6[29],zero,zero
+; AVX512F-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,128,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128]
+; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm1
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm21
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm4
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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]
+; AVX512F-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm4, %ymm3
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm20
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rax), %ymm4
+; AVX512F-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
+; AVX512F-SLOW-NEXT: # ymm0 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm4, %ymm4
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm4, %zmm3
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm7, %ymm2
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[18],zero,zero,zero,zero,ymm7[21],zero,ymm7[19],zero,zero,zero,zero,ymm7[22],zero,ymm7[20]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm8[23],zero,ymm8[21,22,23,26],zero,ymm8[24],zero,ymm8[28,29,26,27]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm8[18,19,20,21],zero,ymm8[19],zero,ymm8[25,26,27,22],zero,ymm8[20],zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm8, %ymm18
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %ymm1, %ymm9, %ymm1
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm12[21],zero,ymm12[19],zero,zero,zero,zero,ymm12[22],zero,ymm12[20],zero,zero
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%rax), %ymm1
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,5,4,u,5,u,4,u,20,21,u,23,u,21,u,23>
+; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm1[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512F-SLOW-NEXT: vpermi2d %zmm0, %zmm2, %zmm25
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm3
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm15
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %xmm2, %xmm15, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm19
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm3, %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm4, %xmm29
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm3, %xmm30
+; AVX512F-SLOW-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm8
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
+; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm0, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm8, %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm8, %xmm22
+; AVX512F-SLOW-NEXT: vpor %xmm0, %xmm2, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = <0,u,0,u,2,3,u,1,u,18,u,19,18,u,19,u>
+; AVX512F-SLOW-NEXT: vmovdqa 32(%rax), %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5,5,6]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpermi2d %zmm0, %zmm2, %zmm8
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %xmm13
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
+; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm0, %xmm8
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm0, %xmm26
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm13, %xmm9
+; AVX512F-SLOW-NEXT: vporq %xmm8, %xmm9, %xmm24
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm27, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm5, %ymm8
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm28, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm6, %ymm9
+; AVX512F-SLOW-NEXT: vpor %ymm8, %ymm9, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm21, %ymm3
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = 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]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm20, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[0,1,14],zero,ymm0[12,13,0,1,14,15],zero,ymm0[3,12,13,2,3,16],zero,ymm0[30,31,28,29,16,17],zero,ymm0[31,18,19,28,29,18],zero
+; AVX512F-SLOW-NEXT: vpor %ymm8, %ymm9, %ymm5
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = zero,zero,zero,zero,zero,zero,ymm11[14],zero,zero,zero,zero,zero,zero,ymm11[15],zero,zero,zero,zero,zero,zero,ymm11[16],zero,zero,zero,zero,zero,zero,ymm11[17],zero,zero,zero,zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm17, %ymm6
+; AVX512F-SLOW-NEXT: vpshufb %ymm6, %ymm10, %ymm9
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm10, %ymm28
+; AVX512F-SLOW-NEXT: vpor %ymm8, %ymm9, %ymm5
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm6
+; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm6, %xmm5
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm6, %xmm20
+; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm9
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm9, %xmm4
+; AVX512F-SLOW-NEXT: vporq %xmm5, %xmm4, %xmm21
+; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm19, %xmm4
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm2, %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm10
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm29, %xmm5
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm10, %xmm7
+; AVX512F-SLOW-NEXT: vporq %xmm4, %xmm7, %xmm19
+; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm5
+; AVX512F-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm5, %xmm4
+; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm7
+; AVX512F-SLOW-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm7, %xmm6
+; AVX512F-SLOW-NEXT: vpor %xmm4, %xmm6, %xmm4
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm0, %zmm4
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = <u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10>
+; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm6, %xmm6
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm5, %xmm27
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm6[0,1,0,1],zmm4[4,5,6,7]
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa (%rax), %xmm12
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm4 = xmm12[0,1,2,3,4,5,5,6]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,3,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
+; AVX512F-SLOW-NEXT: vpandn %ymm4, %ymm11, %ymm4
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = zero,ymm1[13],zero,zero,zero,zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm23
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm18, %ymm4
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm4[30],zero,ymm4[28],zero,zero,zero,zero,ymm4[31],zero,ymm4[29],zero,zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm18
+; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm16, %ymm14
+; AVX512F-SLOW-NEXT: vpshufb %ymm5, %ymm14, %ymm4
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm5, %ymm29
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm8[27],zero,zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm4, %ymm11
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm26, %xmm7
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm13[0],xmm7[0],xmm13[1],xmm7[1],xmm13[2],xmm7[2],xmm13[3],xmm7[3],xmm13[4],xmm7[4],xmm13[5],xmm7[5],xmm13[6],xmm7[6],xmm13[7],xmm7[7]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm11[0,1,2,3],zmm4[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
+; AVX512F-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm0[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm17
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
+; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm4 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
+; AVX512F-SLOW-NEXT: vmovdqa %ymm3, %ymm6
+; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm3, %ymm11
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm11, %zmm26
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm1, %ymm1
+; AVX512F-SLOW-NEXT: vpshuflw $233, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm4 = mem[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,0,1,1,4,4,5,5]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm4, %zmm0
+; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm30, %xmm0
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm0[0],xmm15[0],xmm0[1],xmm15[1],xmm0[2],xmm15[2],xmm0[3],xmm15[3],xmm0[4],xmm15[4],xmm0[5],xmm15[5],xmm0[6],xmm15[6],xmm0[7],xmm15[7]
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm1, %xmm16
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm0[8],xmm15[9],xmm0[9],xmm15[10],xmm0[10],xmm15[11],xmm0[11],xmm15[12],xmm0[12],xmm15[13],xmm0[13],xmm15[14],xmm0[14],xmm15[15],xmm0[15]
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm10[8],xmm2[9],xmm10[9],xmm2[10],xmm10[10],xmm2[11],xmm10[11],xmm2[12],xmm10[12],xmm2[13],xmm10[13],xmm2[14],xmm10[14],xmm2[15],xmm10[15]
+; AVX512F-SLOW-NEXT: vmovdqa %xmm2, %xmm11
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm15 = <6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7>
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm4, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %xmm15, %xmm1, %xmm1
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm0, %zmm30
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm22, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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]
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm20, %xmm5
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm5[8],xmm9[8],xmm5[9],xmm9[9],xmm5[10],xmm9[10],xmm5[11],xmm9[11],xmm5[12],xmm9[12],xmm5[13],xmm9[13],xmm5[14],xmm9[14],xmm5[15],xmm9[15]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u>
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm3, %xmm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm1, %xmm1
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm0, %zmm22
+; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm7[8],xmm13[9],xmm7[9],xmm13[10],xmm7[10],xmm13[11],xmm7[11],xmm13[12],xmm7[12],xmm13[13],xmm7[13],xmm13[14],xmm7[14],xmm13[15],xmm7[15]
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm0
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm24[0,1,0,1],zmm0[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm2 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm4
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm13
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm0[18],zero,zero,zero,zero,ymm0[21],zero,ymm0[19],zero,zero,zero,zero,ymm0[22],zero,ymm0[20]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm31
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm0[23],zero,ymm0[21,22,23,26],zero,ymm0[24],zero,ymm0[28,29,26,27]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm0[18,19,20,21],zero,ymm0[19],zero,ymm0[25,26,27,22],zero,ymm0[20],zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm20
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm6[21],zero,ymm6[19],zero,zero,zero,zero,ymm6[22],zero,ymm6[20],zero,zero
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm28, %ymm1
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm29, %ymm0
+; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm3
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[25],zero,ymm14[23],zero,zero,zero,zero,ymm14[26],zero,ymm14[24],zero,zero
+; AVX512F-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22,128,20,128,18,128,128,128,128,21,128,19,128,128,128,128,22]
+; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm14, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm1, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-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]
+; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm1
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm8, %ymm2
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm2, %ymm29
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128,20,128,18,128,20,21,20,21,128,19,128,19,20,21,22,128]
+; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm8, %ymm8
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm8, %ymm27
+; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512F-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512F-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = zero,ymm0[13],zero,zero,zero,zero,zero,zero,ymm0[14],zero,zero,zero,zero,zero,zero,ymm0[15],zero,zero,zero,zero,zero,zero,ymm0[16],zero,zero,zero,zero,zero,zero,ymm0[17],zero,zero
+; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm14 = ymm0[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[0,1,1,3,4,5,5,7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,3,2]
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm28 = [255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255]
+; AVX512F-SLOW-NEXT: vpandnq %ymm14, %ymm28, %ymm14
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm2, %zmm2
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5>
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm15, %xmm15
+; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm9, %xmm9
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm21, %zmm9, %zmm14
+; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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]
+; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = <4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9>
+; AVX512F-SLOW-NEXT: vmovdqa64 %xmm16, %xmm0
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm0, %xmm8
+; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm9, %xmm9
+; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm19, %zmm9, %zmm9
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm4[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm18[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm3[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm1[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm7[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm6[2,3,2,3]
+; AVX512F-SLOW-NEXT: vmovdqa64 %ymm17, %ymm1
+; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm21 = ymm1[0,0,1,1,4,4,5,5]
+; AVX512F-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512F-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX512F-SLOW-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
+; AVX512F-SLOW-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm1 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm1 = zmm1[0,1,0,1],mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm3 = xmm12[4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm6 = xmm12[1,1,0,0,4,5,6,7]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,1,2,0]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm3
+; AVX512F-SLOW-NEXT: vpandq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm19, %ymm6
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vporq %zmm6, %zmm0, %zmm0
+; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
+; AVX512F-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpand %ymm6, %ymm13, %ymm7
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm7, %zmm5, %zmm5
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vporq %zmm7, %zmm5, %zmm5
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm12 = zmm26[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm12
+; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm12
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm12
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm0 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vporq %zmm0, %zmm7, %zmm0
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm13 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm13
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm5, %zmm13
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm7 = zmm22[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm0, %zmm5, %zmm7
+; AVX512F-SLOW-NEXT: vpternlogq $248, %ymm6, %ymm10, %ymm11
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm21[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpternlogq $236, %ymm6, %ymm4, %ymm0
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm8[0,1,0,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm11, %zmm4
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm5 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpshufhw $190, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm6 = mem[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,2,3,3,6,6,7,7]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm5, %ymm6
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm15[0,1,0,1]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
+; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
+; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm4, %zmm6, %zmm5
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm31[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm20[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpor %ymm4, %ymm8, %ymm4
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm8, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm0
+; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm8 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm10 = mem[0,1,0,1]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm11 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpshuflw $5, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
+; AVX512F-SLOW-NEXT: # xmm15 = mem[1,1,0,0,4,5,6,7]
+; AVX512F-SLOW-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,2,0]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm17 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm18 = ymm29[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm19 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm19 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm27[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm21 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: # ymm22 = mem[2,3,2,3]
+; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm4, %zmm6, %zmm0
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm6 # 32-byte Folded Reload
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm6
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm4, %zmm23
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm23
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm15[0,0,1,0]
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm11, %zmm4
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm4
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm5, %zmm4
+; AVX512F-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload
+; AVX512F-SLOW-NEXT: # zmm5 = mem[2,3,2,3,6,7,6,7]
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm12, %zmm5
+; AVX512F-SLOW-NEXT: vporq %ymm17, %ymm18, %ymm6
+; AVX512F-SLOW-NEXT: vporq %ymm19, %ymm20, %ymm8
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm8[0,1,2,3],zmm6[4,5,6,7]
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm25
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm13, %zmm25
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm24
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm24
+; AVX512F-SLOW-NEXT: vporq %ymm21, %ymm22, %ymm6
+; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6
+; AVX512F-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
+; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm7[0,1,2,3],zmm6[4,5,6,7]
+; AVX512F-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm2
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm2
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm14[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm6 = zmm9[0,1,0,1,4,5,4,5]
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm6
+; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm3[0,0,1,0,4,4,5,4]
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
+; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm0
+; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, 320(%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm24, 256(%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm25, 128(%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm5, 384(%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512F-SLOW-NEXT: vmovdqa64 %zmm23, 64(%rax)
+; AVX512F-SLOW-NEXT: addq $1416, %rsp # imm = 0x588
+; AVX512F-SLOW-NEXT: vzeroupper
+; AVX512F-SLOW-NEXT: retq
;
; AVX512F-ONLY-FAST-LABEL: store_i8_stride7_vf64:
; AVX512F-ONLY-FAST: # %bb.0:
@@ -8115,447 +8268,6 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512F-ONLY-FAST-NEXT: vzeroupper
; AVX512F-ONLY-FAST-NEXT: retq
;
-; AVX512DQ-SLOW-LABEL: store_i8_stride7_vf64:
-; AVX512DQ-SLOW: # %bb.0:
-; AVX512DQ-SLOW-NEXT: subq $1288, %rsp # imm = 0x508
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %ymm1
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,ymm1[14],zero,zero,zero,zero,zero,zero,ymm1[15],zero,zero,zero,zero,zero,zero,ymm1[16],zero,zero,zero,zero,zero,zero,ymm1[17],zero,zero,zero,zero,zero,zero,ymm1[18]
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm1, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %ymm2
-; AVX512DQ-SLOW-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-SLOW-NEXT: vmovdqa %ymm2, %ymm4
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %ymm7
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm1, %ymm21
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %ymm6
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm2, %ymm6, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm16
-; AVX512DQ-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm23
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %ymm2
-; AVX512DQ-SLOW-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-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm3, %ymm18
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm28
-; AVX512DQ-SLOW-NEXT: vporq %ymm0, %ymm1, %ymm22
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %ymm15
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %ymm12
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm12[27],zero,zero,zero,zero,ymm12[30],zero,ymm12[28],zero,zero,zero,zero,ymm12[31],zero,ymm12[29]
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-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-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %ymm3
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[30],zero,ymm14[28],zero,zero,zero,zero,ymm14[31],zero,ymm14[29],zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm1, %ymm3, %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm3, %ymm10
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm8[30],zero,ymm8[28],zero,zero,zero,zero,ymm8[31],zero,ymm8[29],zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-SLOW-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm9, %ymm3
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm9, %ymm25
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rax), %ymm3
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,28,29,26,27,28,29,30,31,30,31,28,29,28,29,30,31]
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,24,25,22,23,24,25,26,27,26,27,24,25]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm7, (%rsp) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm1, %ymm7, %ymm1
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm7[18],zero,zero,zero,zero,ymm7[21],zero,ymm7[19],zero,zero,zero,zero,ymm7[22],zero,ymm7[20]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm6[23],zero,ymm6[21,22,23,26],zero,ymm6[24],zero,ymm6[28,29,26,27]
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm6[18,19,20,21],zero,ymm6[19],zero,ymm6[25,26,27,22],zero,ymm6[20],zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm6, %ymm27
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm4, %ymm0
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm5[21],zero,ymm5[19],zero,zero,zero,zero,ymm5[22],zero,ymm5[20],zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm1, %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm19
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm1, %xmm31
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm2, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm3, %xmm24
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm2, %xmm30
-; AVX512DQ-SLOW-NEXT: vpor %xmm0, %xmm1, %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rdi), %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rsi), %xmm6
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm6, %xmm4
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm0, %xmm5
-; AVX512DQ-SLOW-NEXT: vpor %xmm4, %xmm5, %xmm2
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r9), %xmm11
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%r8), %xmm4
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm11, %xmm5
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm2, %xmm4, %xmm9
-; AVX512DQ-SLOW-NEXT: vpor %xmm5, %xmm9, %xmm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm21, %ymm5
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm5, %ymm10, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm10, %ymm17
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm16, %ymm9
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm9, %ymm14, %ymm9
-; AVX512DQ-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,zero,ymm8[14],zero,zero,zero,zero,zero,zero,ymm8[15],zero,zero,zero,zero,zero,zero,ymm8[16],zero,zero,zero,zero,zero,zero,ymm8[17],zero,zero,zero,zero,zero,zero,ymm8[18]
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm8, %ymm7
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm25, %ymm14
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm14[0,1,14],zero,ymm14[12,13,0,1,14,15],zero,ymm14[3,12,13,2,3,16],zero,ymm14[30,31,28,29,16,17],zero,ymm14[31,18,19,28,29,18],zero
-; AVX512DQ-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm23, %ymm5
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm5, %ymm12, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm18, %ymm8
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm8, %ymm15, %ymm9
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm15, %ymm25
-; AVX512DQ-SLOW-NEXT: vpor %ymm5, %ymm9, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm5, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, %xmm15
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdi), %xmm9
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm9, %xmm3
-; AVX512DQ-SLOW-NEXT: vporq %xmm1, %xmm3, %xmm21
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rcx), %xmm12
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm19, %xmm1
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm12, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rdx), %xmm10
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm3, %xmm10, %xmm3
-; AVX512DQ-SLOW-NEXT: vpor %xmm1, %xmm3, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm5, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%r8), %xmm8
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm2, %xmm8, %xmm3
-; AVX512DQ-SLOW-NEXT: vpor %xmm1, %xmm3, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm5[8],xmm8[9],xmm5[9],xmm8[10],xmm5[10],xmm8[11],xmm5[11],xmm8[12],xmm5[12],xmm8[13],xmm5[13],xmm8[14],xmm5[14],xmm8[15],xmm5[15]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm2, %xmm3, %xmm3
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm2, %xmm16
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,0,1],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rax), %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,5,6]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255]
-; AVX512DQ-SLOW-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa (%rax), %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = zero,ymm2[13],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
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm22
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm27, %ymm1
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm1, %ymm23
-; AVX512DQ-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm28, %ymm2
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm2, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm3, %ymm29
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm3 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm13[27],zero,zero,zero,zero,ymm13[30],zero,ymm13[28],zero,zero,zero,zero,ymm13[31],zero,ymm13[29]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm3
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7]
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm3[0,1,2,3],zmm1[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm14[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm14, %ymm24
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,3,3,6,6,7,7]
-; AVX512DQ-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10,9,8,7,0,0,0,11,10]
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm7, %ymm14
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm7, %ymm7
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm7, %zmm26
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm20, %ymm1
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,ymm1[30],zero,ymm1[28],zero,zero,zero,zero,ymm1[31],zero,ymm1[29],zero,zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm5, %ymm20
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm3, %ymm1, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm3 = ymm8[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,1,1,4,4,5,5]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm27
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm30, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm31, %xmm3
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-SLOW-NEXT: vmovdqa64 %xmm5, %xmm19
-; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
-; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = <6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm7, %xmm3, %xmm3
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm7, %xmm1, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm3, %zmm30
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
-; AVX512DQ-SLOW-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-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm15[8],xmm9[8],xmm15[9],xmm9[9],xmm15[10],xmm9[10],xmm15[11],xmm9[11],xmm15[12],xmm9[12],xmm15[13],xmm9[13],xmm15[14],xmm9[14],xmm15[15],xmm9[15]
-; AVX512DQ-SLOW-NEXT: vmovdqa %xmm15, %xmm5
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm6 = <2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm6, %xmm3, %xmm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm1, %ymm28
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm6, %xmm0, %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm31
-; AVX512DQ-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm11[8],xmm4[9],xmm11[9],xmm4[10],xmm11[10],xmm4[11],xmm11[11],xmm4[12],xmm11[12],xmm4[13],xmm11[13],xmm4[14],xmm11[14],xmm4[15],xmm11[15]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm16, %xmm1
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm1, %xmm0, %xmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,0,1],zmm0[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512DQ-SLOW-NEXT: vmovdqu (%rsp), %ymm1 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm1, %ymm1
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm17, %ymm11
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm11, %ymm6
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[18],zero,zero,zero,zero,ymm11[21],zero,ymm11[19],zero,zero,zero,zero,ymm11[22],zero,ymm11[20]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25],zero,ymm15[23],zero,ymm15[21,22,23,26],zero,ymm15[24],zero,ymm15[28,29,26,27]
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,ymm15[18,19,20,21],zero,ymm15[19],zero,ymm15[25,26,27,22],zero,ymm15[20],zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm0, %ymm16
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm14[21],zero,ymm14[19],zero,zero,zero,zero,ymm14[22],zero,ymm14[20],zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm0, %ymm17
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm25, %ymm14
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm29, %ymm0
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm0, %ymm14, %ymm3
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm2, %ymm0
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,ymm2[25],zero,ymm2[23],zero,zero,zero,zero,ymm2[26],zero,ymm2[24],zero,zero
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm0, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm14, %ymm0
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm15, %ymm0
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm13, %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm2, %ymm29
-; AVX512DQ-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm13, %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vpshufb %ymm4, %ymm15, %ymm2
-; AVX512DQ-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = zero,ymm2[13],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
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm2[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm15 = ymm15[0,1,1,3,4,5,5,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,3,2]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255,0,255,255,255,255,255,255]
-; AVX512DQ-SLOW-NEXT: vpandnq %ymm15, %ymm25, %ymm15
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm4, %zmm4
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5>
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm14, %xmm7, %xmm15
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm14, %xmm9, %xmm5
-; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, %xmm21, %zmm5, %zmm25
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm21 = ymm1[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm23[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm12[0],xmm10[1],xmm12[1],xmm10[2],xmm12[2],xmm10[3],xmm12[3],xmm10[4],xmm12[4],xmm10[5],xmm12[5],xmm10[6],xmm12[6],xmm10[7],xmm12[7]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9>
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %xmm19, %xmm2
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm2, %xmm12
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} ymm5 = ymm8[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[2,2,3,3,6,6,7,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm3[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm11[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm18[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm16[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm17[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %ymm24, %ymm7
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} ymm7 = ymm7[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,0,1,1,4,4,5,5]
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm1, %xmm8
-; AVX512DQ-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm1
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vporq %zmm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm19 = [18374966859431673855,18446463693966278655,18374966859431673855,18446463693966278655]
-; AVX512DQ-SLOW-NEXT: # ymm19 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpandq %ymm19, %ymm6, %ymm0
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vporq %zmm2, %zmm0, %zmm2
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm3 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [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-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm3, %zmm0
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm1 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm2 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vporq %zmm1, %zmm2, %zmm2
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm6 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm27[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm1
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm2, %zmm3, %zmm1
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm6 = zmm30[0,1,0,1,4,5,4,5]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm31[0,1,0,1,4,5,4,5]
-; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm6, %zmm3, %zmm2
-; AVX512DQ-SLOW-NEXT: vpternlogq $248, %ymm19, %ymm21, %ymm14
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm7[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpternlogq $236, %ymm19, %ymm10, %ymm3
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm14, %zmm6
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm9[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpternlogq $236, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm20, %ymm7
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
-; AVX512DQ-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255,255,255,255,0,0,255,255]
-; AVX512DQ-SLOW-NEXT: vpternlogq $184, %zmm6, %zmm9, %zmm7
-; AVX512DQ-SLOW-NEXT: vpor %ymm5, %ymm11, %ymm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm6, %zmm3
-; AVX512DQ-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm3
-; AVX512DQ-SLOW-NEXT: vinserti32x4 $2, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm19 # 16-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm28[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vpermq $68, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm8 = mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm9 = ymm5[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-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # xmm10 = xmm10[0],mem[0],xmm10[1],mem[1],xmm10[2],mem[2],xmm10[3],mem[3],xmm10[4],mem[4],xmm10[5],mem[5],xmm10[6],mem[6],xmm10[7],mem[7]
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512DQ-SLOW-NEXT: vmovdqa 32(%rax), %xmm11
-; AVX512DQ-SLOW-NEXT: vshufi64x2 $0, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm21 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm21 = zmm10[0,1,0,1],mem[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm12 = xmm10[1,1,0,0,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,1,2,0]
-; AVX512DQ-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,4,5,4,5,8,9,6,7,6,7,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm10, %xmm14
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm12, %zmm12
-; AVX512DQ-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm11[1,1,0,0,4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[0,1,2,0]
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm15 = mem[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm29[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm17 = mem[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm18 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm18 = mem[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vmovdqa %ymm5, %ymm10
-; AVX512DQ-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[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-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[2,1,1,2,4,5,6,7,10,9,9,10,12,13,14,15]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,1,1,3,4,5,5,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpshufb %xmm13, %xmm11, %xmm13
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,1,0]
-; AVX512DQ-SLOW-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,5,5,6]
-; AVX512DQ-SLOW-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,3,3]
-; AVX512DQ-SLOW-NEXT: vpermq $238, (%rsp), %ymm23 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm23 = mem[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # ymm26 = mem[2,3,2,3]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm8
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm6, %zmm22
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm22
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm14[0,0,1,0]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm8, %zmm6
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm7, %zmm6
-; AVX512DQ-SLOW-NEXT: vpermq $238, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Folded Reload
-; AVX512DQ-SLOW-NEXT: # zmm7 = mem[2,3,2,3,6,7,6,7]
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm7
-; AVX512DQ-SLOW-NEXT: vporq %ymm15, %ymm16, %ymm0
-; AVX512DQ-SLOW-NEXT: vporq %ymm17, %ymm18, %ymm8
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm8[0,1,2,3],zmm0[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm10[2,2,3,2]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm8, %zmm5
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm5
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm5
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm11[0,1,0,1]
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm13, %zmm0
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm0
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
-; AVX512DQ-SLOW-NEXT: vporq %ymm23, %ymm26, %ymm1
-; AVX512DQ-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
-; AVX512DQ-SLOW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm2[0,1,2,3],zmm1[4,5,6,7]
-; AVX512DQ-SLOW-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm4
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm25[0,1,0,1,4,5,4,5]
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm19[0,1,0,1,4,5,4,5]
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm2
-; AVX512DQ-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm12[0,0,1,0,4,4,5,4]
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm21, %zmm1
-; AVX512DQ-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm1
-; AVX512DQ-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm4, 320(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm0, 256(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm5, 128(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm7, 384(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-SLOW-NEXT: vmovdqa64 %zmm22, 64(%rax)
-; AVX512DQ-SLOW-NEXT: addq $1288, %rsp # imm = 0x508
-; AVX512DQ-SLOW-NEXT: vzeroupper
-; AVX512DQ-SLOW-NEXT: retq
-;
; AVX512DQ-FAST-LABEL: store_i8_stride7_vf64:
; AVX512DQ-FAST: # %bb.0:
; AVX512DQ-FAST-NEXT: subq $1496, %rsp # imm = 0x5D8
@@ -9008,320 +8720,319 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-FAST-NEXT: vzeroupper
; AVX512DQ-FAST-NEXT: retq
;
-; AVX512BW-SLOW-LABEL: store_i8_stride7_vf64:
-; AVX512BW-SLOW: # %bb.0:
-; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-SLOW-NEXT: vmovdqa (%rax), %ymm13
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rax), %ymm9
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm1, %ymm13, %ymm0
-; AVX512BW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
-; AVX512BW-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vpermw %ymm13, %ymm2, %ymm2
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %ymm11
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm2 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm2, %ymm11, %ymm3
-; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %ymm12
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm4, %ymm12, %ymm5
-; AVX512BW-SLOW-NEXT: vpor %ymm3, %ymm5, %ymm3
-; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %xmm6
-; AVX512BW-SLOW-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm5
-; AVX512BW-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm5, %zmm10
-; AVX512BW-SLOW-NEXT: movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
-; AVX512BW-SLOW-NEXT: kmovq %r10, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm10 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdx), %ymm18
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm0, %ymm18, %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rcx), %ymm17
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm3, %ymm17, %ymm7
-; AVX512BW-SLOW-NEXT: vpor %ymm5, %ymm7, %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdx), %xmm29
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rcx), %xmm23
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm23[8],xmm29[8],xmm23[9],xmm29[9],xmm23[10],xmm29[10],xmm23[11],xmm29[11],xmm23[12],xmm29[12],xmm23[13],xmm29[13],xmm23[14],xmm29[14],xmm23[15],xmm29[15]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm19
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdi), %ymm21
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm21, %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rsi), %ymm22
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm26, %ymm22, %ymm7
-; AVX512BW-SLOW-NEXT: vpor %ymm5, %ymm7, %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdi), %xmm20
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm14
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rsi), %xmm16
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rsi), %xmm15
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm16[8],xmm20[8],xmm16[9],xmm20[9],xmm16[10],xmm20[10],xmm16[11],xmm20[11],xmm16[12],xmm20[12],xmm16[13],xmm20[13],xmm16[14],xmm20[14],xmm16[15],xmm20[15]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm25 = xmm25[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm25, %zmm7
-; AVX512BW-SLOW-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
-; AVX512BW-SLOW-NEXT: kmovq %r10, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm19, %zmm7 {%k1}
-; AVX512BW-SLOW-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
-; AVX512BW-SLOW-NEXT: kmovq %r10, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm10, %zmm7 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
-; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm10, %ymm10
-; AVX512BW-SLOW-NEXT: vpshufb %ymm1, %ymm9, %ymm1
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %ymm28
-; AVX512BW-SLOW-NEXT: vpshufb %ymm2, %ymm28, %ymm2
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm31
-; AVX512BW-SLOW-NEXT: vpshufb %ymm4, %ymm31, %ymm4
-; AVX512BW-SLOW-NEXT: vpor %ymm2, %ymm4, %ymm2
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm31[20],zero,ymm31[18],zero,ymm31[20,21,20,21],zero,ymm31[19],zero,ymm31[19,20,21,22],zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm28[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm28[20],zero,ymm28[18],zero,zero,zero,zero,ymm28[21],zero,ymm28[19],zero,zero,zero,zero,ymm28[22]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpor %ymm4, %ymm10, %ymm4
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm4
-; AVX512BW-SLOW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
-; AVX512BW-SLOW-NEXT: kmovq %r10, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm4 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rsi), %ymm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm25
-; AVX512BW-SLOW-NEXT: vpshuflw {{.*#+}} ymm2 = ymm25[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,1,1,4,4,5,5]
-; AVX512BW-SLOW-NEXT: movl $676341840, %esi # imm = 0x28502850
-; AVX512BW-SLOW-NEXT: kmovd %esi, %k1
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm2 {%k1} = ymm5[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm25, %ymm10
-; AVX512BW-SLOW-NEXT: vpshufb %ymm26, %ymm5, %ymm19
-; AVX512BW-SLOW-NEXT: vporq %ymm10, %ymm19, %ymm10
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm2, %zmm10, %zmm10
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdx), %ymm2
-; AVX512BW-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm19
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rcx), %ymm1
-; AVX512BW-SLOW-NEXT: vpshufb %ymm3, %ymm1, %ymm3
-; AVX512BW-SLOW-NEXT: vporq %ymm19, %ymm3, %ymm3
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm19 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[18],zero,ymm2[18,19,20,21],zero,ymm2[19],zero,ymm2[25,26,27,22],zero,ymm2[20],zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm19[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm24 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm1[18],zero,zero,zero,zero,ymm1[21],zero,ymm1[19],zero,zero,zero,zero,ymm1[22],zero,ymm1[20]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vporq %ymm19, %ymm24, %ymm19
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm3, %zmm3
-; AVX512BW-SLOW-NEXT: movabsq $3485998880071096368, %rsi # imm = 0x3060C183060C1830
-; AVX512BW-SLOW-NEXT: kmovq %rsi, %k3
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm3, %zmm10 {%k3}
-; AVX512BW-SLOW-NEXT: movabsq $-4357498600088870461, %rsi # imm = 0xC3870E1C3870E1C3
-; AVX512BW-SLOW-NEXT: kmovq %rsi, %k2
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm4, %zmm10 {%k2}
-; AVX512BW-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm21[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,2,3,3,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpbroadcastd {{.*#+}} ymm6 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
-; AVX512BW-SLOW-NEXT: movl $338170920, %esi # imm = 0x14281428
-; AVX512BW-SLOW-NEXT: kmovd %esi, %k2
-; AVX512BW-SLOW-NEXT: vpshufb %ymm6, %ymm22, %ymm3 {%k2}
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3],xmm14[4],xmm15[4],xmm14[5],xmm15[5],xmm14[6],xmm15[6],xmm14[7],xmm15[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm26
-; AVX512BW-SLOW-NEXT: vpshufhw {{.*#+}} ymm3 = ymm18[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,3,3,4,6,7,7]
-; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm8, %ymm17, %ymm4
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm3, %ymm4 {%k1}
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm4[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm4
-; AVX512BW-SLOW-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm27 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm27 = xmm27[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm27, %zmm19, %zmm19
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm26, %zmm19 {%k3}
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm26 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm11[27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm26 = ymm26[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm27 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm12[27],zero,zero,zero,zero,ymm12[30],zero,ymm12[28],zero,zero,zero,zero,ymm12[31],zero,ymm12[29]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vporq %ymm26, %ymm27, %ymm30
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r9), %xmm26
-; AVX512BW-SLOW-NEXT: vmovdqa64 32(%r8), %xmm27
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm30, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rax), %zmm30
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm24 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
-; AVX512BW-SLOW-NEXT: vpermw %zmm30, %zmm24, %zmm24
-; AVX512BW-SLOW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k3
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm24, %zmm0 {%k3}
-; AVX512BW-SLOW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k3
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm19 {%k3}
-; AVX512BW-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm25[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,2,3,3,6,6,7,7]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm6, %ymm5, %ymm0 {%k2}
-; AVX512BW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
-; AVX512BW-SLOW-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm6, %ymm5, %ymm5
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm25, %ymm25
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vporq %ymm5, %ymm25, %ymm5
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm5, %zmm25
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[24,25],zero,ymm2[23],zero,ymm2[21,22,23,26],zero,ymm2[24],zero,ymm2[28,29,26,27]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm5 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm1[25],zero,ymm1[23],zero,zero,zero,zero,ymm1[26],zero,ymm1[24],zero,zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpor %ymm0, %ymm5, %ymm0
-; AVX512BW-SLOW-NEXT: vpshufb %ymm8, %ymm1, %ymm1
-; AVX512BW-SLOW-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,3,3,4,6,7,7]
-; AVX512BW-SLOW-NEXT: vmovdqu8 %ymm2, %ymm1 {%k1}
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: movabsq $1742999440035548184, %rax # imm = 0x183060C183060C18
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k2
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm25 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%r9), %zmm0
-; AVX512BW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm31[0,1,2,3],zmm0[4,5,6,7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm1[23],zero,zmm1[23,24,25,26],zero,zmm1[24],zero,zmm1[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm1[59],zero,zero,zero,zero,zmm1[62],zero,zmm1[60],zero,zero,zero,zero,zmm1[63],zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%r8), %zmm2
-; AVX512BW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm28[0,1,2,3],zmm2[4,5,6,7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm5[25],zero,zmm5[23],zero,zero,zero,zero,zmm5[26],zero,zmm5[24],zero,zero,zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm5[59],zero,zero,zero,zero,zmm5[62],zero,zmm5[60],zero,zero,zero,zero,zmm5[63],zero,zmm5[61]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vporq %zmm1, %zmm5, %zmm1
-; AVX512BW-SLOW-NEXT: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k2
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm25 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
-; AVX512BW-SLOW-NEXT: vpermi2w %zmm30, %zmm9, %zmm1
-; AVX512BW-SLOW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k3
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm25 {%k3}
-; AVX512BW-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm21[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
-; AVX512BW-SLOW-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,1,1,4,4,5,5]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} ymm1 {%k1} = ymm22[u,u,u,u,5,u,3,u,u,u,u,6,u,4,u,u,u,u,u,u,21,u,19,u,u,u,u,22,u,20,u,u]
-; AVX512BW-SLOW-NEXT: vpshufb %ymm6, %ymm22, %ymm5
-; AVX512BW-SLOW-NEXT: vpshufb %ymm24, %ymm21, %ymm6
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vpor %ymm5, %ymm6, %ymm5
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,3,2,3]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm1, %zmm1
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rdx), %zmm5
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm5, %zmm5
-; AVX512BW-SLOW-NEXT: vmovdqa64 (%rcx), %zmm6
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm6, %zmm6
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm5 = zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm5[18,19,20,21],zero,zmm5[19],zero,zmm5[25,26,27,22],zero,zmm5[20],zero,zmm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm5[55],zero,zmm5[53,54,55,58],zero,zmm5[56],zero,zmm5[60,61,58,59]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm6 = zmm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm6[18],zero,zero,zero,zero,zmm6[21],zero,zmm6[19],zero,zero,zero,zero,zmm6[22],zero,zmm6[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm6[57],zero,zmm6[55],zero,zero,zero,zero,zmm6[58],zero,zmm6[56],zero,zero,zero,zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vporq %zmm5, %zmm6, %zmm17
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm17 {%k2}
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm2, %zmm1
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm1[18],zero,zmm1[20,21,20,21],zero,zmm1[19],zero,zmm1[19,20,21,22],zero,zmm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm1[55],zero,zmm1[55,56,57,58],zero,zmm1[56],zero,zmm1[62,63]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm0[20],zero,zmm0[18],zero,zero,zero,zero,zmm0[21],zero,zmm0[19],zero,zero,zero,zero,zmm0[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm0[57],zero,zmm0[55],zero,zero,zero,zero,zmm0[58],zero,zmm0[56],zero,zero
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,3,2,3,6,7,6,7]
-; AVX512BW-SLOW-NEXT: vporq %zmm1, %zmm0, %zmm1
-; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm30, %zmm0
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
-; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm2, %zmm2
-; AVX512BW-SLOW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm2, %zmm1 {%k1}
-; AVX512BW-SLOW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm17 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm2, %xmm3, %xmm1
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm5, %xmm4, %xmm6
-; AVX512BW-SLOW-NEXT: vpor %xmm1, %xmm6, %xmm1
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm3, %zmm1, %zmm1
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm3, %xmm15, %xmm4
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm6 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm6, %xmm14, %xmm8
-; AVX512BW-SLOW-NEXT: vpor %xmm4, %xmm8, %xmm4
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm1 = xmm8[2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm1, %zmm4, %zmm1
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm11, %zmm1 {%k2}
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm4, %xmm26, %xmm8
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm11 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
-; AVX512BW-SLOW-NEXT: vpermi2w %zmm30, %zmm9, %zmm11
-; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} xmm9 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
-; AVX512BW-SLOW-NEXT: vpshufb %xmm9, %xmm27, %xmm12
-; AVX512BW-SLOW-NEXT: vpor %xmm8, %xmm12, %xmm8
-; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm12 = xmm12[u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm12, %zmm8, %zmm8
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm11, %zmm8 {%k1}
-; AVX512BW-SLOW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm8, %zmm1 {%k1}
-; AVX512BW-SLOW-NEXT: vpshufb %xmm2, %xmm23, %xmm2
-; AVX512BW-SLOW-NEXT: vpshufb %xmm5, %xmm29, %xmm5
-; AVX512BW-SLOW-NEXT: vpor %xmm2, %xmm5, %xmm2
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm29[0],xmm23[0],xmm29[1],xmm23[1],xmm29[2],xmm23[2],xmm29[3],xmm23[3],xmm29[4],xmm23[4],xmm29[5],xmm23[5],xmm29[6],xmm23[6],xmm29[7],xmm23[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm5, %zmm2
-; AVX512BW-SLOW-NEXT: vpshufb %xmm3, %xmm16, %xmm3
-; AVX512BW-SLOW-NEXT: vpshufb %xmm6, %xmm20, %xmm5
-; AVX512BW-SLOW-NEXT: vpor %xmm3, %xmm5, %xmm3
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm20[0],xmm16[0],xmm20[1],xmm16[1],xmm20[2],xmm16[2],xmm20[3],xmm16[3],xmm20[4],xmm16[4],xmm20[5],xmm16[5],xmm20[6],xmm16[6],xmm20[7],xmm16[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm3, %zmm5, %zmm3
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm3 = zmm3[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512BW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
-; AVX512BW-SLOW-NEXT: vpshufb %xmm4, %xmm5, %xmm2
-; AVX512BW-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512BW-SLOW-NEXT: vpshufb %xmm9, %xmm6, %xmm4
-; AVX512BW-SLOW-NEXT: vpor %xmm2, %xmm4, %xmm2
-; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
-; AVX512BW-SLOW-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u]
-; AVX512BW-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm4, %zmm2
-; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm4 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
-; AVX512BW-SLOW-NEXT: vpermw %zmm0, %zmm4, %zmm0
-; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,1,0,1,4,5,4,5]
-; AVX512BW-SLOW-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm2 {%k1}
-; AVX512BW-SLOW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
-; AVX512BW-SLOW-NEXT: kmovq %rax, %k1
-; AVX512BW-SLOW-NEXT: vmovdqu8 %zmm2, %zmm3 {%k1}
-; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm3, (%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm10, 320(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm1, 256(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm17, 128(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm25, 384(%rax)
-; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm19, 192(%rax)
-; AVX512BW-SLOW-NEXT: vzeroupper
-; AVX512BW-SLOW-NEXT: retq
+; AVX512BW-ONLY-SLOW-LABEL: store_i8_stride7_vf64:
+; AVX512BW-ONLY-SLOW: # %bb.0:
+; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rax), %ymm15
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm15, %ymm0
+; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512BW-ONLY-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm15, %ymm1, %ymm1
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r9), %ymm10
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm20, %ymm10, %ymm0
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %ymm11
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm11, %ymm1
+; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm0
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm1
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm4, %xmm6, %xmm6
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm21
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm3, %zmm21 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdx), %ymm16
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [0,1,0,1,14,128,14,15,0,1,14,15,128,13,14,15,16,17,16,128,30,31,30,31,16,17,128,31,28,29,30,31]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm16, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rcx), %ymm17
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm23, %ymm17, %ymm5
+; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm3, %ymm5, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm7, %xmm7
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm22
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %ymm18
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [0,1,14,128,12,13,0,1,14,15,128,3,12,13,2,3,16,128,30,31,28,29,16,17,128,31,18,19,28,29,18,128]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm18, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rsi), %ymm19
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm26, %ymm19, %ymm7
+; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm8
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm27 = 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-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm27, %xmm27
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm27, %zmm3
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm22, %zmm3 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm21, %zmm3 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm2, %ymm21, %ymm21
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm9, %ymm2, %ymm9
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm9, %zmm9
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r9), %ymm21
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm20, %ymm21, %ymm20
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r8), %ymm22
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm22, %ymm24
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm20, %ymm24, %ymm20
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm24 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm22[20],zero,ymm22[18],zero,ymm22[20,21,20,21],zero,ymm22[19],zero,ymm22[19,20,21,22],zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm27 = ymm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm21[20],zero,ymm21[18],zero,zero,zero,zero,ymm21[21],zero,ymm21[19],zero,zero,zero,zero,ymm21[22]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm24, %ymm27, %ymm24
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm24
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm9, %zmm24 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm27
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm28
+; AVX512BW-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm28[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,1,1,4,4,5,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastd {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
+; AVX512BW-ONLY-SLOW-NEXT: movl $676341840, %r10d # imm = 0x28502850
+; AVX512BW-ONLY-SLOW-NEXT: kmovd %r10d, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm20, %ymm27, %ymm9 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm28, %ymm25
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm26, %ymm27, %ymm26
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm25, %ymm26, %ymm25
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm25, %zmm9
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm26
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm14, %ymm26, %ymm14
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm29
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm23, %ymm29, %ymm23
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm14, %ymm23, %ymm14
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm23 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm26[18],zero,ymm26[18,19,20,21],zero,ymm26[19],zero,ymm26[25,26,27,22],zero,ymm26[20],zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm25 = ymm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm29[18],zero,zero,zero,zero,ymm29[21],zero,ymm29[19],zero,zero,zero,zero,ymm29[22],zero,ymm29[20]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm23, %ymm25, %ymm23
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm14, %zmm14
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k2
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm14, %zmm9 {%k2}
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm24, %zmm9 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm28[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,3,3,6,6,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastd {{.*#+}} ymm24 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
+; AVX512BW-ONLY-SLOW-NEXT: movl $338170920, %r10d # imm = 0x14281428
+; AVX512BW-ONLY-SLOW-NEXT: kmovd %r10d, %k4
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm27, %ymm14 {%k4}
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm23 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512BW-ONLY-SLOW-NEXT: # ymm23 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm23, %ymm27, %ymm25
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm28, %ymm28
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm25, %ymm28, %ymm25
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm25, %zmm14
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm25 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm26[24,25],zero,ymm26[23],zero,ymm26[21,22,23,26],zero,ymm26[24],zero,ymm26[28,29,26,27]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm28 = ymm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm29[25],zero,ymm29[23],zero,zero,zero,zero,ymm29[26],zero,ymm29[24],zero,zero,zero,zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm25, %ymm28, %ymm28
+; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm25 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm29, %ymm29
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm26 = ymm26[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm26 = ymm26[0,2,3,3,4,6,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm26, %ymm29 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdx), %zmm26
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm29[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm28
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r8), %zmm29
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm28, %zmm14 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r9), %zmm28
+; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm28[4,5,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm22[23],zero,zmm22[23,24,25,26],zero,zmm22[24],zero,zmm22[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[0,1,2,3],zmm29[4,5,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm21 = zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm21[25],zero,zmm21[23],zero,zero,zero,zero,zmm21[26],zero,zmm21[24],zero,zero,zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm21[59],zero,zero,zero,zero,zmm21[62],zero,zmm21[60],zero,zero,zero,zero,zmm21[63],zero,zmm21[61]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm21 = zmm21[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %zmm22, %zmm21, %zmm22
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rax), %zmm21
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k3
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm22, %zmm14 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm22 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512BW-ONLY-SLOW-NEXT: vpermi2w %zmm21, %zmm2, %zmm22
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm22, %zmm14 {%k5}
+; AVX512BW-ONLY-SLOW-NEXT: vpshuflw {{.*#+}} ymm22 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm30 = ymm22[0,0,1,1,4,4,5,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm20, %ymm19, %ymm30 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rcx), %zmm20
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm23, %ymm19, %ymm22
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm18, %ymm23
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %ymm22, %ymm23, %ymm23
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm30[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm27, %zmm27
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r8), %xmm23
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm26, %zmm26
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm26 = zmm26[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm26[18,19,20,21],zero,zmm26[19],zero,zmm26[25,26,27,22],zero,zmm26[20],zero,zmm26[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm26[55],zero,zmm26[53,54,55,58],zero,zmm26[56],zero,zmm26[60,61,58,59]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm26 = zmm26[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm20, %zmm20
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm20 = zmm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm20[18],zero,zero,zero,zero,zmm20[21],zero,zmm20[19],zero,zero,zero,zero,zmm20[22],zero,zmm20[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm20[57],zero,zmm20[55],zero,zero,zero,zero,zmm20[58],zero,zmm20[56],zero,zero,zero,zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %zmm26, %zmm20, %zmm20
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdx), %xmm26
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm27, %zmm20 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rcx), %xmm27
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm29, %zmm29
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm29 = zmm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm29[18],zero,zmm29[20,21,20,21],zero,zmm29[19],zero,zmm29[19,20,21,22],zero,zmm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm29[55],zero,zmm29[55,56,57,58],zero,zmm29[56],zero,zmm29[62,63]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm28, %zmm28
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} zmm28 = zmm28[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm28[20],zero,zmm28[18],zero,zero,zero,zero,zmm28[21],zero,zmm28[19],zero,zero,zero,zero,zmm28[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm28[57],zero,zmm28[55],zero,zero,zero,zero,zmm28[58],zero,zmm28[56],zero,zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
+; AVX512BW-ONLY-SLOW-NEXT: vporq %zmm29, %zmm28, %zmm29
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm15
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
+; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm15, %zmm28, %zmm28
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm28, %zmm29 {%k5}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdi), %xmm28
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k5
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm29, %zmm20 {%k5}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rsi), %xmm29
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm18 = ymm18[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm30 = ymm18[2,2,3,3,6,6,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm19, %ymm30 {%k4}
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm28[0],xmm29[0],xmm28[1],xmm29[1],xmm28[2],xmm29[2],xmm28[3],xmm29[3],xmm28[4],xmm29[4],xmm28[5],xmm29[5],xmm28[6],xmm29[6],xmm28[7],xmm29[7]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm18 = <0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm18, %xmm19, %xmm19
+; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm30[2,3,2,3],zmm19[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm24 = xmm26[0],xmm27[0],xmm26[1],xmm27[1],xmm26[2],xmm27[2],xmm26[3],xmm27[3],xmm26[4],xmm27[4],xmm26[5],xmm27[5],xmm26[6],xmm27[6],xmm26[7],xmm27[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm25, %ymm17, %ymm25
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm17 = <4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm17, %xmm24, %xmm24
+; AVX512BW-ONLY-SLOW-NEXT: vpshufhw {{.*#+}} ymm16 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} ymm16 = ymm16[0,2,3,3,4,6,7,7]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %ymm16, %ymm25 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm25[2,3,2,3],zmm24[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm19, %zmm16 {%k2}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm10[27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512BW-ONLY-SLOW-NEXT: vpor %ymm10, %ymm11, %ymm11
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm19 = 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-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = <u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm19, %xmm19
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm19[0,1,0,1]
+; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm11, %zmm11
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm19 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm21, %zmm19, %zmm19
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm19, %zmm11 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm11, %zmm16 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm19 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm19, %xmm27, %xmm11
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm24 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm24, %xmm26, %xmm25
+; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm11, %xmm25, %xmm11
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm25, %xmm12
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm12, %zmm11, %zmm11
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm29, %xmm25
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm26 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm26, %xmm28, %xmm27
+; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm25, %xmm27, %xmm25
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm27 = 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-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm28 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm27, %xmm11
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm11, %zmm25, %zmm11
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm28, %zmm11 {%k3}
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm22, %xmm25
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm27 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512BW-ONLY-SLOW-NEXT: vpermi2w %zmm21, %zmm2, %zmm27
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm2, %xmm23, %xmm21
+; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm25, %xmm21, %xmm21
+; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm22 = 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-ONLY-SLOW-NEXT: vpshufb %xmm4, %xmm22, %xmm4
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm21, %zmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm27, %zmm4 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm4, %zmm11 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm19, %xmm6, %xmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm24, %xmm5, %xmm19
+; AVX512BW-ONLY-SLOW-NEXT: vporq %xmm4, %xmm19, %xmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm17, %xmm5, %xmm5
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm8, %xmm5
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm26, %xmm7, %xmm6
+; AVX512BW-ONLY-SLOW-NEXT: vpor %xmm5, %xmm6, %xmm5
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm18, %xmm6, %xmm6
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm5, %zmm6, %zmm5
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm4, %zmm5 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm13, %xmm0, %xmm4
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm2, %xmm1, %xmm2
+; AVX512BW-ONLY-SLOW-NEXT: vpor %xmm4, %xmm2, %xmm2
+; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm0
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
+; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm15, %zmm1, %zmm1
+; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
+; AVX512BW-ONLY-SLOW-NEXT: kmovq %rax, %k1
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm0, %zmm5 {%k1}
+; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm11, 256(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, 192(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, 128(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512BW-ONLY-SLOW-NEXT: vzeroupper
+; AVX512BW-ONLY-SLOW-NEXT: retq
;
; AVX512BW-FAST-LABEL: store_i8_stride7_vf64:
; AVX512BW-FAST: # %bb.0:
@@ -9637,6 +9348,320 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FAST-NEXT: addq $200, %rsp
; AVX512BW-FAST-NEXT: vzeroupper
; AVX512BW-FAST-NEXT: retq
+;
+; AVX512DQBW-SLOW-LABEL: store_i8_stride7_vf64:
+; AVX512DQBW-SLOW: # %bb.0:
+; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rax), %ymm15
+; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rax), %ymm2
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [12,13,2,3,12,13,0,1,14,15,2,3,0,1,14,15,28,29,18,19,28,29,16,17,30,31,18,19,16,17,30,31]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm9, %ymm15, %ymm0
+; AVX512DQBW-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [4,5,4,5,5,6,5,6,4,5,4,5,5,6,5,6]
+; AVX512DQBW-SLOW-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vpermw %ymm15, %ymm1, %ymm1
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm3
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%r9), %ymm10
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [13,128,128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm20, %ymm10, %ymm0
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %ymm11
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm24 = [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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm11, %ymm1
+; AVX512DQBW-SLOW-NEXT: vpor %ymm0, %ymm1, %ymm5
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%r9), %xmm0
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %xmm1
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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]
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,6,7,2,3,u,u,u,8,9,4,5,u,u,u,10>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm4, %xmm6, %xmm6
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm21
+; AVX512DQBW-SLOW-NEXT: movabsq $2323999253380730912, %r10 # imm = 0x2040810204081020
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm3, %zmm21 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdx), %ymm16
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm14, %ymm16, %ymm3
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rcx), %ymm17
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [128,128,128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm23, %ymm17, %ymm5
+; AVX512DQBW-SLOW-NEXT: vpor %ymm3, %ymm5, %ymm3
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdx), %xmm5
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <6,3,2,u,u,u,9,8,5,4,u,u,u,11,10,7>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm7, %xmm7
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm7, %zmm22
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %ymm18
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm25 = [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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm18, %ymm3
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rsi), %ymm19
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [128,128,128,14,128,128,128,128,128,128,15,128,128,128,128,128,128,16,128,128,128,128,128,128,17,128,128,128,128,128,128,18]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm26, %ymm19, %ymm7
+; AVX512DQBW-SLOW-NEXT: vpor %ymm3, %ymm7, %ymm3
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdi), %xmm7
+; AVX512DQBW-SLOW-NEXT: vmovdqa (%rsi), %xmm8
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm27 = 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]
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <2,u,u,u,9,8,5,4,u,u,u,11,10,7,6,u>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm13, %xmm27, %xmm27
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm27, %zmm3
+; AVX512DQBW-SLOW-NEXT: movabsq $435749860008887046, %r10 # imm = 0x60C183060C18306
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm22, %zmm3 {%k1}
+; AVX512DQBW-SLOW-NEXT: movabsq $4066998693416279096, %r10 # imm = 0x3870E1C3870E1C38
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm21, %zmm3 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm21 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10]
+; AVX512DQBW-SLOW-NEXT: vpermw %ymm2, %ymm21, %ymm21
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm9, %ymm2, %ymm9
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm9, %zmm9
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r9), %ymm21
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm20, %ymm21, %ymm20
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm22
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm22, %ymm24
+; AVX512DQBW-SLOW-NEXT: vporq %ymm20, %ymm24, %ymm20
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm24 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm22[20],zero,ymm22[18],zero,ymm22[20,21,20,21],zero,ymm22[19],zero,ymm22[19,20,21,22],zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm27 = ymm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm21[20],zero,ymm21[18],zero,zero,zero,zero,ymm21[21],zero,ymm21[19],zero,zero,zero,zero,ymm21[22]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vporq %ymm24, %ymm27, %ymm24
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm24
+; AVX512DQBW-SLOW-NEXT: movabsq $145249953336295682, %r10 # imm = 0x204081020408102
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm9, %zmm24 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm27
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm28
+; AVX512DQBW-SLOW-NEXT: vpshuflw {{.*#+}} ymm9 = ymm28[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,1,1,4,4,5,5]
+; AVX512DQBW-SLOW-NEXT: vpbroadcastd {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6,5,4,3,6]
+; AVX512DQBW-SLOW-NEXT: movl $676341840, %r10d # imm = 0x28502850
+; AVX512DQBW-SLOW-NEXT: kmovd %r10d, %k1
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm20, %ymm27, %ymm9 {%k1}
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm28, %ymm25
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm26, %ymm27, %ymm26
+; AVX512DQBW-SLOW-NEXT: vporq %ymm25, %ymm26, %ymm25
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm25, %zmm9
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm26
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm14, %ymm26, %ymm14
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm29
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm23, %ymm29, %ymm23
+; AVX512DQBW-SLOW-NEXT: vporq %ymm14, %ymm23, %ymm14
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm23 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm26[18],zero,ymm26[18,19,20,21],zero,ymm26[19],zero,ymm26[25,26,27,22],zero,ymm26[20],zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm25 = ymm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm29[18],zero,zero,zero,zero,ymm29[21],zero,ymm29[19],zero,zero,zero,zero,ymm29[22],zero,ymm29[20]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vporq %ymm23, %ymm25, %ymm23
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm14, %zmm14
+; AVX512DQBW-SLOW-NEXT: movabsq $3485998880071096368, %r10 # imm = 0x3060C183060C1830
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k2
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm14, %zmm9 {%k2}
+; AVX512DQBW-SLOW-NEXT: movabsq $-4357498600088870461, %r10 # imm = 0xC3870E1C3870E1C3
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k3
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm24, %zmm9 {%k3}
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm14 = ymm28[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm14 = ymm14[2,2,3,3,6,6,7,7]
+; AVX512DQBW-SLOW-NEXT: vpbroadcastd {{.*#+}} ymm24 = [13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14,13,12,15,14]
+; AVX512DQBW-SLOW-NEXT: movl $338170920, %r10d # imm = 0x14281428
+; AVX512DQBW-SLOW-NEXT: kmovd %r10d, %k4
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm27, %ymm14 {%k4}
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm23 = [25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128]
+; AVX512DQBW-SLOW-NEXT: # ymm23 = mem[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm23, %ymm27, %ymm25
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm27 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,23,128,128,128,128,26,128,24,128,128,128,128,27,128,25]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm28, %ymm28
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vporq %ymm25, %ymm28, %ymm25
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm25, %zmm14
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm25 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm26[24,25],zero,ymm26[23],zero,ymm26[21,22,23,26],zero,ymm26[24],zero,ymm26[28,29,26,27]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm25 = ymm25[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm28 = ymm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,ymm29[25],zero,ymm29[23],zero,zero,zero,zero,ymm29[26],zero,ymm29[24],zero,zero,zero,zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vporq %ymm25, %ymm28, %ymm28
+; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm25 = [11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12,11,0,0,0,15,14,13,12]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm29, %ymm29
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm26 = ymm26[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm26 = ymm26[0,2,3,3,4,6,7,7]
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm26, %ymm29 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdx), %zmm26
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm29[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm28
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r8), %zmm29
+; AVX512DQBW-SLOW-NEXT: movabsq $1742999440035548184, %r10 # imm = 0x183060C183060C18
+; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k3
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm28, %zmm14 {%k3}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r9), %zmm28
+; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm22 = zmm22[0,1,2,3],zmm28[4,5,6,7]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm22 = zmm22[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25],zero,zmm22[23],zero,zmm22[23,24,25,26],zero,zmm22[24],zero,zmm22[30,31,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,61],zero,zmm22[59],zero,zero,zero,zero,zmm22[62],zero,zmm22[60],zero,zero,zero,zero,zmm22[63],zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm22 = zmm22[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm21 = zmm21[0,1,2,3],zmm29[4,5,6,7]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm21 = zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm21[25],zero,zmm21[23],zero,zero,zero,zero,zmm21[26],zero,zmm21[24],zero,zero,zmm21[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm21[59],zero,zero,zero,zero,zmm21[62],zero,zmm21[60],zero,zero,zero,zero,zmm21[63],zero,zmm21[61]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm21 = zmm21[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vporq %zmm22, %zmm21, %zmm22
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rax), %zmm21
+; AVX512DQBW-SLOW-NEXT: movabsq $6971997760142192736, %rax # imm = 0x60C183060C183060
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k3
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm22, %zmm14 {%k3}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm22 = [11,13,12,11,12,13,13,12,11,13,12,11,12,13,13,12,62,61,62,63,63,62,62,63,62,61,62,63,63,62,62,63]
+; AVX512DQBW-SLOW-NEXT: vpermi2w %zmm21, %zmm2, %zmm22
+; AVX512DQBW-SLOW-NEXT: movabsq $-9150747060186627967, %rax # imm = 0x8102040810204081
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k5
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm22, %zmm14 {%k5}
+; AVX512DQBW-SLOW-NEXT: vpshuflw {{.*#+}} ymm22 = ymm18[1,2,2,3,4,5,6,7,9,10,10,11,12,13,14,15]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm30 = ymm22[0,0,1,1,4,4,5,5]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm20, %ymm19, %ymm30 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rcx), %zmm20
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm23, %ymm19, %ymm22
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm18, %ymm23
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vporq %ymm22, %ymm23, %ymm23
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r9), %xmm22
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm30[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm23, %zmm27, %zmm27
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r8), %xmm23
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm26, %zmm26
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm26 = zmm26[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18],zero,zmm26[18,19,20,21],zero,zmm26[19],zero,zmm26[25,26,27,22],zero,zmm26[20],zero,zmm26[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57],zero,zmm26[55],zero,zmm26[53,54,55,58],zero,zmm26[56],zero,zmm26[60,61,58,59]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm26 = zmm26[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm20, %zmm20
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm20 = zmm20[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm20[18],zero,zero,zero,zero,zmm20[21],zero,zmm20[19],zero,zero,zero,zero,zmm20[22],zero,zmm20[20,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zmm20[57],zero,zmm20[55],zero,zero,zero,zero,zmm20[58],zero,zmm20[56],zero,zero,zero,zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm20 = zmm20[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vporq %zmm26, %zmm20, %zmm20
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdx), %xmm26
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm27, %zmm20 {%k3}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rcx), %xmm27
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm29, %zmm29
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm29 = zmm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20],zero,zmm29[18],zero,zmm29[20,21,20,21],zero,zmm29[19],zero,zmm29[19,20,21,22],zero,zmm29[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,56,57,56,57],zero,zmm29[55],zero,zmm29[55,56,57,58],zero,zmm29[56],zero,zmm29[62,63]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm29 = zmm29[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm28, %zmm28
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} zmm28 = zmm28[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zmm28[20],zero,zmm28[18],zero,zero,zero,zero,zmm28[21],zero,zmm28[19],zero,zero,zero,zero,zmm28[22,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,zero,zero,zero,zmm28[57],zero,zmm28[55],zero,zero,zero,zero,zmm28[58],zero,zmm28[56],zero,zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm28 = zmm28[2,3,2,3,6,7,6,7]
+; AVX512DQBW-SLOW-NEXT: vporq %zmm29, %zmm28, %zmm29
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm21, %zmm15
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = [10,9,9,10,10,9,9,10,9,10,14,15,10,9,9,10,27,29,28,27,28,29,29,28,27,29,28,27,28,29,29,28]
+; AVX512DQBW-SLOW-NEXT: vpermw %zmm15, %zmm28, %zmm28
+; AVX512DQBW-SLOW-NEXT: movabsq $1161999626690365456, %rax # imm = 0x1020408102040810
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k5
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm28, %zmm29 {%k5}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdi), %xmm28
+; AVX512DQBW-SLOW-NEXT: movabsq $2033499346708139548, %rax # imm = 0x1C3870E1C3870E1C
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k5
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm29, %zmm20 {%k5}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rsi), %xmm29
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm18 = ymm18[0,1,2,3,6,7,7,6,8,9,10,11,14,15,15,14]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm30 = ymm18[2,2,3,3,6,6,7,7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm19, %ymm30 {%k4}
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm28[0],xmm29[0],xmm28[1],xmm29[1],xmm28[2],xmm29[2],xmm28[3],xmm29[3],xmm28[4],xmm29[4],xmm28[5],xmm29[5],xmm28[6],xmm29[6],xmm28[7],xmm29[7]
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm18 = <0,1,u,u,u,6,7,2,3,u,u,u,8,9,4,5>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm18, %xmm19, %xmm19
+; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm30[2,3,2,3],zmm19[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm24 = xmm26[0],xmm27[0],xmm26[1],xmm27[1],xmm26[2],xmm27[2],xmm26[3],xmm27[3],xmm26[4],xmm27[4],xmm26[5],xmm27[5],xmm26[6],xmm27[6],xmm26[7],xmm27[7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %ymm25, %ymm17, %ymm25
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm17 = <4,5,0,1,u,u,u,6,7,2,3,u,u,u,8,9>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm17, %xmm24, %xmm24
+; AVX512DQBW-SLOW-NEXT: vpshufhw {{.*#+}} ymm16 = ymm16[0,1,2,3,7,6,6,7,8,9,10,11,15,14,14,15]
+; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} ymm16 = ymm16[0,2,3,3,4,6,7,7]
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %ymm16, %ymm25 {%k1}
+; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm16 = zmm25[2,3,2,3],zmm24[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm19, %zmm16 {%k2}
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,29],zero,ymm10[27],zero,zero,zero,zero,ymm10[30],zero,ymm10[28],zero,zero,zero,zero,ymm10[31],zero
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u],zero,ymm11[27],zero,zero,zero,zero,ymm11[30],zero,ymm11[28],zero,zero,zero,zero,ymm11[31],zero,ymm11[29]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,3,2,3]
+; AVX512DQBW-SLOW-NEXT: vpor %ymm10, %ymm11, %ymm11
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm19 = 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]
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = <u,u,4,5,0,1,u,u,u,6,7,2,3,u,u,u>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm10, %xmm19, %xmm19
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm19 = ymm19[0,1,0,1]
+; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm19, %zmm11, %zmm11
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm19 = [14,13,14,15,15,14,14,15,14,13,14,15,15,14,14,15,17,17,16,16,17,17,16,16,20,21,17,17,17,17,16,16]
+; AVX512DQBW-SLOW-NEXT: vpermw %zmm21, %zmm19, %zmm19
+; AVX512DQBW-SLOW-NEXT: movabsq $580999813345182728, %rax # imm = 0x810204081020408
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm19, %zmm11 {%k1}
+; AVX512DQBW-SLOW-NEXT: movabsq $1016749673354069774, %rax # imm = 0xE1C3870E1C3870E
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm11, %zmm16 {%k1}
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm19 = <u,u,u,128,7,128,5,u,u,u,128,8,128,6,u,u>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm19, %xmm27, %xmm11
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm24 = <u,u,u,7,128,5,128,u,u,u,8,128,6,128,u,u>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm24, %xmm26, %xmm25
+; AVX512DQBW-SLOW-NEXT: vporq %xmm11, %xmm25, %xmm11
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm25 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm25, %xmm12
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm12, %zmm11, %zmm11
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <u,128,7,128,5,u,u,u,128,8,128,6,u,u,u,128>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm29, %xmm25
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} xmm26 = <u,7,128,5,128,u,u,u,8,128,6,128,u,u,u,9>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm26, %xmm28, %xmm27
+; AVX512DQBW-SLOW-NEXT: vporq %xmm25, %xmm27, %xmm25
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm27 = 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]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm28 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm13, %xmm27, %xmm11
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm11, %zmm25, %zmm11
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm11 = zmm11[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm28, %zmm11 {%k3}
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <128,4,u,u,u,128,7,128,5,u,u,u,128,8,128,6>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm13, %xmm22, %xmm25
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm27 = [2,2,2,4,2,2,2,4,3,3,3,3,2,2,2,4,52,53,52,53,53,54,53,54,52,53,52,53,53,54,53,54]
+; AVX512DQBW-SLOW-NEXT: vpermi2w %zmm21, %zmm2, %zmm27
+; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <4,128,u,u,u,7,128,5,128,u,u,u,8,128,6,128>
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm2, %xmm23, %xmm21
+; AVX512DQBW-SLOW-NEXT: vporq %xmm25, %xmm21, %xmm21
+; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm22 = 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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm4, %xmm22, %xmm4
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm21, %zmm4
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: movabsq $290499906672591364, %rax # imm = 0x408102040810204
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm27, %zmm4 {%k1}
+; AVX512DQBW-SLOW-NEXT: movabsq $-8714997200177740921, %rax # imm = 0x870E1C3870E1C387
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm4, %zmm11 {%k1}
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm19, %xmm6, %xmm4
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm24, %xmm5, %xmm19
+; AVX512DQBW-SLOW-NEXT: vporq %xmm4, %xmm19, %xmm4
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm17, %xmm5, %xmm5
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm8, %xmm5
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm26, %xmm7, %xmm6
+; AVX512DQBW-SLOW-NEXT: vpor %xmm5, %xmm6, %xmm5
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm18, %xmm6, %xmm6
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm5, %zmm6, %zmm5
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm4 = zmm4[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: movabsq $871499720017774092, %rax # imm = 0xC183060C183060C
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm4, %zmm5 {%k1}
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm13, %xmm0, %xmm4
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm2, %xmm1, %xmm2
+; AVX512DQBW-SLOW-NEXT: vpor %xmm4, %xmm2, %xmm2
+; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; AVX512DQBW-SLOW-NEXT: vpshufb %xmm10, %xmm0, %xmm0
+; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm0, %zmm0
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [1,1,0,0,1,1,0,0,4,5,1,1,1,1,0,0,18,18,18,20,18,18,18,20,19,19,19,19,18,18,18,20]
+; AVX512DQBW-SLOW-NEXT: vpermw %zmm15, %zmm1, %zmm1
+; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,1,0,1,4,5,4,5]
+; AVX512DQBW-SLOW-NEXT: movabsq $4647998506761461824, %rax # imm = 0x4081020408102040
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
+; AVX512DQBW-SLOW-NEXT: movabsq $8133997386832558192, %rax # imm = 0x70E1C3870E1C3870
+; AVX512DQBW-SLOW-NEXT: kmovq %rax, %k1
+; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm0, %zmm5 {%k1}
+; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, 320(%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm11, 256(%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm16, 192(%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm20, 128(%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm3, 64(%rax)
+; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512DQBW-SLOW-NEXT: vzeroupper
+; AVX512DQBW-SLOW-NEXT: retq
%in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <64 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <64 x i8>, ptr %in.vecptr2, align 64
@@ -9663,9 +9688,7 @@ define void @store_i8_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512-FAST: {{.*}}
; AVX512-SLOW: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
-; AVX512BW-ONLY-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
-; AVX512DQBW-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
More information about the llvm-commits
mailing list