[llvm] [X86] combineTargetShuffle - fold vpermv3(widen(x),mask,widen(y)) -> vpermv(widen(concat(x,y)),mask') (PR #203031)
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Wed Jun 10 08:55:56 PDT 2026
https://github.com/RKSimon created https://github.com/llvm/llvm-project/pull/203031
We already handle the case where the src vectors were half size, but we can generalize this to widening from xmm to zmm as well - mainly to help non-VLX builds
A couple of codesize increases in non-VLX builds - mostly from additional asm / kill comments, but also due to a couple of poor folds in combineConcatVectorOps that need further yak shaving.
>From dbe3d058f9def445e54bc7793c28b795aacafd6b Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Wed, 10 Jun 2026 15:53:51 +0100
Subject: [PATCH] [X86] combineTargetShuffle - fold
vpermv3(widen(x),mask,widen(y)) -> vpermv(widen(concat(x,y)),mask')
We already handle the case where the src vectors were half size, but we can generalize this to widening from xmm to zmm as well - mainly to help non-VLX builds
A couple of codesize increases in non-VLX builds - mostly from asm comments / kill comments, but also due to a couple of poor folds in combineConcatVectorOps that need further yak shaving.
---
llvm/lib/Target/X86/X86ISelLowering.cpp | 20 +-
llvm/test/CodeGen/X86/vector-fshr-128.ll | 16 +-
llvm/test/CodeGen/X86/vector-fshr-rot-128.ll | 13 +-
.../vector-interleaved-load-i8-stride-8.ll | 2578 +++++------
.../vector-interleaved-store-i16-stride-4.ll | 172 +-
.../vector-interleaved-store-i16-stride-7.ll | 3874 +++++++++--------
.../vector-interleaved-store-i8-stride-5.ll | 630 +--
.../vector-interleaved-store-i8-stride-8.ll | 380 +-
8 files changed, 3705 insertions(+), 3978 deletions(-)
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index eb65d5d6dcc66..899b7645b4c14 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -43974,20 +43974,26 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
return lowerShuffleWithPERMV(DL, VT, Mask, N.getOperand(0),
DAG.getUNDEF(VT), Subtarget, DAG);
}
- // If sources are half width, then concat and use VPERMV with adjusted
- // mask.
+ // If sources are widened, then concat and use VPERMV with adjusted mask.
SDValue Ops[2];
- MVT HalfVT = VT.getHalfNumVectorElementsVT();
if (sd_match(V1,
m_InsertSubvector(m_Undef(), m_Value(Ops[0]), m_Zero())) &&
sd_match(V2,
m_InsertSubvector(m_Undef(), m_Value(Ops[1]), m_Zero())) &&
- Ops[0].getValueType() == HalfVT && Ops[1].getValueType() == HalfVT) {
+ (Ops[0].getValueSizeInBits() % VT.getScalarSizeInBits()) == 0 &&
+ Ops[0].getValueType() == Ops[1].getValueType() &&
+ Ops[0].getValueType().isSimple()) {
+ MVT SubVT = Ops[0].getSimpleValueType();
+ MVT ConcatVT = SubVT.getDoubleNumVectorElementsVT();
+ unsigned NumSubElts = SubVT.getSizeInBits() / VT.getScalarSizeInBits();
if (SDValue ConcatSrc =
- combineConcatVectorOps(DL, VT, Ops, DAG, Subtarget)) {
+ combineConcatVectorOps(DL, ConcatVT, Ops, DAG, Subtarget)) {
+ ConcatSrc = widenSubVector(ConcatSrc, false, Subtarget, DAG, DL,
+ VT.getSizeInBits());
for (int &M : Mask)
- M = (M < (int)NumElts ? M : (M - (NumElts / 2)));
- return lowerShuffleWithPERMV(DL, VT, Mask, ConcatSrc,
+ M = (M < (int)NumElts ? M : (M - (NumElts - NumSubElts)));
+ return lowerShuffleWithPERMV(DL, VT, Mask,
+ DAG.getBitcast(VT, ConcatSrc),
DAG.getUNDEF(VT), Subtarget, DAG);
}
}
diff --git a/llvm/test/CodeGen/X86/vector-fshr-128.ll b/llvm/test/CodeGen/X86/vector-fshr-128.ll
index 2ef640e99945d..f29afd9f3af0e 100644
--- a/llvm/test/CodeGen/X86/vector-fshr-128.ll
+++ b/llvm/test/CodeGen/X86/vector-fshr-128.ll
@@ -1546,14 +1546,14 @@ define <16 x i8> @splatvar_funnnel_v16i8(<16 x i8> %x, <16 x i8> %y, <16 x i8> %
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v16i8:
; AVX512VBMI2: # %bb.0:
-; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm3 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78]
-; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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]
-; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
-; AVX512VBMI2-NEXT: vpsrlw %xmm2, %xmm4, %xmm4
-; AVX512VBMI2-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]
-; AVX512VBMI2-NEXT: vpsrlw %xmm2, %xmm0, %xmm0
-; AVX512VBMI2-NEXT: vpermt2b %zmm4, %zmm3, %zmm0
-; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
+; AVX512VBMI2-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
+; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} ymm3 = [0,64,1,65,2,66,3,67,4,68,5,69,6,70,7,71,8,72,9,73,10,74,11,75,12,76,13,77,14,78,15,79]
+; AVX512VBMI2-NEXT: vpermi2b %zmm0, %zmm1, %zmm3
+; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm0
+; AVX512VBMI2-NEXT: vpsrlw %xmm0, %ymm3, %ymm0
+; AVX512VBMI2-NEXT: vpmovwb %zmm0, %ymm0
+; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
; AVX512VBMI2-NEXT: vzeroupper
; AVX512VBMI2-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll b/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
index 313cb18e36f3b..7ecfdbeacac8c 100644
--- a/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
+++ b/llvm/test/CodeGen/X86/vector-fshr-rot-128.ll
@@ -1209,14 +1209,13 @@ define <16 x i8> @splatvar_funnnel_v16i8(<16 x i8> %x, <16 x i8> %amt) nounwind
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v16i8:
; AVX512VBMI2: # %bb.0:
-; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78]
+; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX512VBMI2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
-; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX512VBMI2-NEXT: vpsrlw %xmm1, %xmm3, %xmm3
-; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; AVX512VBMI2-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
-; AVX512VBMI2-NEXT: vpermt2b %zmm3, %zmm2, %zmm0
-; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
+; AVX512VBMI2-NEXT: vpsrlw %xmm1, %ymm0, %ymm0
+; AVX512VBMI2-NEXT: vpmovwb %zmm0, %ymm0
+; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm0
; AVX512VBMI2-NEXT: vzeroupper
; AVX512VBMI2-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
index 1271a0f5f52e8..6c0970f9d67de 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll
@@ -6852,253 +6852,207 @@ define void @load_i8_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm19
-; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm2
-; AVX512BW-NEXT: vmovdqa 240(%rdi), %xmm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm1, %xmm4
-; AVX512BW-NEXT: vmovdqa 224(%rdi), %xmm3
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm3, %xmm5
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512BW-NEXT: vmovdqa 208(%rdi), %xmm4
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm4, %xmm6
-; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm7
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm7, %xmm9
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [2,11,2,11,2,11,2,11]
-; AVX512BW-NEXT: vpermt2d %ymm5, %ymm17, %ymm6
-; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm5
-; AVX512BW-NEXT: vpmovqb %ymm5, %xmm9
+; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0
+; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm1
+; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm2
+; AVX512BW-NEXT: vpmovqb %zmm2, %xmm2
+; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm3
+; AVX512BW-NEXT: vpmovqb %ymm3, %xmm5
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm16 = [0,9,0,9,0,9,0,9]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm9
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3,4,5],ymm6[6,7]
-; AVX512BW-NEXT: vmovdqa 112(%rdi), %xmm6
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm6, %xmm10
-; AVX512BW-NEXT: vmovdqa 96(%rdi), %xmm9
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm9, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
-; AVX512BW-NEXT: vmovdqa 80(%rdi), %xmm10
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm10, %xmm13
-; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm12
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm13[0],xmm11[1],xmm13[1],xmm11[2],xmm13[2],xmm11[3],xmm13[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1,2],xmm8[3]
-; AVX512BW-NEXT: vpmovqb %zmm19, %xmm11
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1],xmm8[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512BW-NEXT: vmovdqa 160(%rdi), %xmm13
-; AVX512BW-NEXT: vmovdqa 176(%rdi), %xmm14
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm1, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm3, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm4, %xmm15
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm7, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512BW-NEXT: vpermt2d %ymm8, %ymm17, %ymm15
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm14, %xmm18
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm13, %xmm8
+; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm5
+; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [18,2,27,11,18,2,27,11,18,2,27,11,18,2,27,11]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm1[2,3,6,7,0,1,4,5]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm2
+; AVX512BW-NEXT: vpermw %zmm2, %zmm17, %zmm2
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512BW-NEXT: vmovdqa 112(%rdi), %xmm5
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm5, %xmm8
+; AVX512BW-NEXT: vmovdqa 96(%rdi), %xmm6
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm6, %xmm7
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
+; AVX512BW-NEXT: vmovdqa 80(%rdi), %xmm7
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm10
+; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm9
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm8[3]
+; AVX512BW-NEXT: vpmovqb %zmm0, %xmm8
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512BW-NEXT: vmovdqa 160(%rdi), %xmm10
+; AVX512BW-NEXT: vmovdqa 176(%rdi), %xmm11
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $8, %ymm3, %ymm2
+; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm12
+; AVX512BW-NEXT: vpermw %zmm12, %zmm17, %zmm12
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm12[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm12 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm5, %xmm13
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm6, %xmm12
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm13
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm12[3]
+; AVX512BW-NEXT: vpsrlq $8, %zmm0, %zmm12
+; AVX512BW-NEXT: vpmovqb %zmm12, %xmm12
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm12[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $16, %ymm3, %ymm2
+; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm13
+; AVX512BW-NEXT: vpermw %zmm13, %zmm17, %zmm13
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm13[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm13 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm13, %xmm5, %xmm14
+; AVX512BW-NEXT: vpshufb %xmm13, %xmm6, %xmm13
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm14
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1],xmm1[2],xmm14[2],xmm1[3],xmm14[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm13[3]
+; AVX512BW-NEXT: vpsrlq $16, %zmm0, %zmm13
+; AVX512BW-NEXT: vpmovqb %zmm13, %xmm13
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $24, %ymm3, %ymm2
+; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm14
+; AVX512BW-NEXT: vpermw %zmm14, %zmm17, %zmm14
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm14[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm14 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm14, %xmm5, %xmm15
+; AVX512BW-NEXT: vpshufb %xmm14, %xmm6, %xmm14
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm15
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm14[3]
+; AVX512BW-NEXT: vpsrlq $24, %zmm0, %zmm14
+; AVX512BW-NEXT: vpmovqb %zmm14, %xmm14
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $32, %ymm3, %ymm2
+; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm15
+; AVX512BW-NEXT: vpermw %zmm15, %zmm17, %zmm15
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm15[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm15 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm15, %xmm5, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm15, %xmm6, %xmm15
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm18[0],xmm15[1],xmm18[1],xmm15[2],xmm18[2],xmm15[3],xmm18[3]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm18[0],xmm1[1],xmm18[1],xmm1[2],xmm18[2],xmm1[3],xmm18[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm15[3]
+; AVX512BW-NEXT: vpsrlq $32, %zmm0, %zmm15
+; AVX512BW-NEXT: vpmovqb %zmm15, %xmm15
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $40, %ymm3, %ymm2
+; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm4, %zmm18
+; AVX512BW-NEXT: vpermw %zmm18, %zmm17, %zmm8
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm8[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm5, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm6, %xmm8
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm18[0],xmm8[1],xmm18[1],xmm8[2],xmm18[2],xmm8[3],xmm18[3]
-; AVX512BW-NEXT: vpsrlq $8, %ymm5, %ymm18
-; AVX512BW-NEXT: vpmovqb %ymm18, %xmm0
-; AVX512BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm0
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm15[6,7]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm6, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm9, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm10, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1,2],xmm2[3]
-; AVX512BW-NEXT: vpsrlq $8, %zmm19, %zmm8
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm7, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm18[0],xmm1[1],xmm18[1],xmm1[2],xmm18[2],xmm1[3],xmm18[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm8[3]
+; AVX512BW-NEXT: vpsrlq $40, %zmm0, %zmm8
; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1],xmm2[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm7, %xmm15
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm14, %xmm15
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
-; AVX512BW-NEXT: vpsrlq $16, %ymm5, %ymm15
-; AVX512BW-NEXT: vpmovqb %ymm15, %xmm15
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $16, %zmm19, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm7, %xmm15
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm14, %xmm15
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
-; AVX512BW-NEXT: vpsrlq $24, %ymm5, %ymm15
-; AVX512BW-NEXT: vpmovqb %ymm15, %xmm15
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $24, %zmm19, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm7, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm18[0],xmm11[0],xmm18[1],xmm11[1],xmm18[2],xmm11[2],xmm18[3],xmm11[3]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm14, %xmm18
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm18[0],xmm2[1],xmm18[1],xmm2[2],xmm18[2],xmm2[3],xmm18[3]
-; AVX512BW-NEXT: vpsrlq $32, %ymm5, %ymm18
-; AVX512BW-NEXT: vpmovqb %ymm18, %xmm15
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $32, %zmm19, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm4, %xmm15
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm7, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm15
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm14, %xmm18
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm11, %xmm8
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm10, %xmm2
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512BW-NEXT: vpsrlq $48, %ymm3, %ymm8
+; AVX512BW-NEXT: vpmovqb %ymm8, %xmm8
+; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm8
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %zmm2, %zmm4, %zmm18
+; AVX512BW-NEXT: vpermw %zmm18, %zmm17, %zmm12
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm5, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm12, %xmm6, %xmm12
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm18[0],xmm12[1],xmm18[1],xmm12[2],xmm18[2],xmm12[3],xmm18[3]
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm7, %xmm18
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm9, %xmm2
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm18[0],xmm2[1],xmm18[1],xmm2[2],xmm18[2],xmm2[3],xmm18[3]
-; AVX512BW-NEXT: vpsrlq $40, %ymm5, %ymm18
-; AVX512BW-NEXT: vpmovqb %ymm18, %xmm11
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm11
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3,4,5],ymm15[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1,2],xmm12[3]
+; AVX512BW-NEXT: vpsrlq $48, %zmm0, %zmm12
+; AVX512BW-NEXT: vpmovqb %zmm12, %xmm12
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm12[0,1],xmm2[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm11, %xmm11
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm10, %xmm8
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $40, %zmm19, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm4, %xmm15
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm7, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512BW-NEXT: vpermt2d %ymm8, %ymm17, %ymm15
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm14, %xmm18
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm13, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm18[0],xmm8[1],xmm18[1],xmm8[2],xmm18[2],xmm8[3],xmm18[3]
-; AVX512BW-NEXT: vpsrlq $48, %ymm5, %ymm18
-; AVX512BW-NEXT: vpmovqb %ymm18, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm15[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm10, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $48, %zmm19, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm3, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm4, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm7, %xmm4
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512BW-NEXT: vpermt2d %ymm1, %ymm17, %ymm3
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm14, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
-; AVX512BW-NEXT: vpsrlq $56, %ymm5, %ymm4
-; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
-; AVX512BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm4
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm6, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm10, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $56, %zmm19, %zmm2
-; AVX512BW-NEXT: vpmovqb %zmm2, %xmm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
-; AVX512BW-NEXT: vmovdqa64 %ymm20, (%rsi)
-; AVX512BW-NEXT: vmovdqa64 %ymm21, (%rdx)
-; AVX512BW-NEXT: vmovdqa64 %ymm22, (%rcx)
-; AVX512BW-NEXT: vmovdqa64 %ymm23, (%r8)
-; AVX512BW-NEXT: vmovdqa64 %ymm24, (%r9)
-; AVX512BW-NEXT: vmovdqa64 %ymm25, (%r11)
-; AVX512BW-NEXT: vmovdqa %ymm8, (%r10)
+; AVX512BW-NEXT: vpsrlq $56, %ymm3, %ymm3
+; AVX512BW-NEXT: vpmovqb %ymm3, %xmm3
+; AVX512BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm3
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %zmm8, %zmm4, %zmm4
+; AVX512BW-NEXT: vpermw %zmm4, %zmm17, %zmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm5, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm6, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm7, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm8, %xmm9, %xmm6
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512BW-NEXT: vpsrlq $56, %zmm0, %zmm0
+; AVX512BW-NEXT: vpmovqb %zmm0, %xmm0
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX512BW-NEXT: vmovdqa64 %ymm19, (%rsi)
+; AVX512BW-NEXT: vmovdqa64 %ymm20, (%rdx)
+; AVX512BW-NEXT: vmovdqa %ymm13, (%rcx)
+; AVX512BW-NEXT: vmovdqa %ymm14, (%r8)
+; AVX512BW-NEXT: vmovdqa %ymm15, (%r9)
+; AVX512BW-NEXT: vmovdqa %ymm1, (%r11)
+; AVX512BW-NEXT: vmovdqa %ymm2, (%r10)
; AVX512BW-NEXT: vmovdqa %ymm0, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
@@ -7276,253 +7230,207 @@ define void @load_i8_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm19
-; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa 240(%rdi), %xmm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm1, %xmm4
-; AVX512DQ-BW-NEXT: vmovdqa 224(%rdi), %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm3, %xmm5
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-BW-NEXT: vmovdqa 208(%rdi), %xmm4
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm4, %xmm6
-; AVX512DQ-BW-NEXT: vmovdqa 192(%rdi), %xmm7
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm7, %xmm9
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
-; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [2,11,2,11,2,11,2,11]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm5, %ymm17, %ymm6
-; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %ymm5
-; AVX512DQ-BW-NEXT: vpmovqb %ymm5, %xmm9
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm0
+; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm1
+; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm2
+; AVX512DQ-BW-NEXT: vpmovqb %zmm2, %xmm2
+; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %ymm3
+; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm5
; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm16 = [0,9,0,9,0,9,0,9]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm9
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm9[0,1,2,3,4,5],ymm6[6,7]
-; AVX512DQ-BW-NEXT: vmovdqa 112(%rdi), %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm6, %xmm10
-; AVX512DQ-BW-NEXT: vmovdqa 96(%rdi), %xmm9
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm9, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
-; AVX512DQ-BW-NEXT: vmovdqa 80(%rdi), %xmm10
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm10, %xmm13
-; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm12
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm13[0],xmm11[1],xmm13[1],xmm11[2],xmm13[2],xmm11[3],xmm13[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1,2],xmm8[3]
-; AVX512DQ-BW-NEXT: vpmovqb %zmm19, %xmm11
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1],xmm8[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm20
-; AVX512DQ-BW-NEXT: vmovdqa 160(%rdi), %xmm13
-; AVX512DQ-BW-NEXT: vmovdqa 176(%rdi), %xmm14
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm1, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm3, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm4, %xmm15
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm7, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm17, %ymm15
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm14, %xmm18
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm13, %xmm8
+; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm5
+; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm17 = [18,2,27,11,18,2,27,11,18,2,27,11,18,2,27,11]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm1[2,3,6,7,0,1,4,5]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm2
+; AVX512DQ-BW-NEXT: vpermw %zmm2, %zmm17, %zmm2
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512DQ-BW-NEXT: vmovdqa 112(%rdi), %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm5, %xmm8
+; AVX512DQ-BW-NEXT: vmovdqa 96(%rdi), %xmm6
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm6, %xmm7
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
+; AVX512DQ-BW-NEXT: vmovdqa 80(%rdi), %xmm7
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm10
+; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm9
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm8[3]
+; AVX512DQ-BW-NEXT: vpmovqb %zmm0, %xmm8
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqa64 %ymm1, %ymm19
+; AVX512DQ-BW-NEXT: vmovdqa 160(%rdi), %xmm10
+; AVX512DQ-BW-NEXT: vmovdqa 176(%rdi), %xmm11
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm3, %ymm2
+; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm12
+; AVX512DQ-BW-NEXT: vpermw %zmm12, %zmm17, %zmm12
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm12[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm12 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm5, %xmm13
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm6, %xmm12
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm13
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1],xmm1[2],xmm13[2],xmm1[3],xmm13[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm12[3]
+; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm0, %zmm12
+; AVX512DQ-BW-NEXT: vpmovqb %zmm12, %xmm12
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm12[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqa64 %ymm1, %ymm20
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm3, %ymm2
+; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm13
+; AVX512DQ-BW-NEXT: vpermw %zmm13, %zmm17, %zmm13
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm13[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm13 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm13, %xmm5, %xmm14
+; AVX512DQ-BW-NEXT: vpshufb %xmm13, %xmm6, %xmm13
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm14
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1],xmm1[2],xmm14[2],xmm1[3],xmm14[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm13[3]
+; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm0, %zmm13
+; AVX512DQ-BW-NEXT: vpmovqb %zmm13, %xmm13
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm3, %ymm2
+; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm14
+; AVX512DQ-BW-NEXT: vpermw %zmm14, %zmm17, %zmm14
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm14[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm14 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm14, %xmm5, %xmm15
+; AVX512DQ-BW-NEXT: vpshufb %xmm14, %xmm6, %xmm14
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm15
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm14[3]
+; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm0, %zmm14
+; AVX512DQ-BW-NEXT: vpmovqb %zmm14, %xmm14
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm3, %ymm2
+; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm15
+; AVX512DQ-BW-NEXT: vpermw %zmm15, %zmm17, %zmm15
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm15[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm15 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm5, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm6, %xmm15
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm18[0],xmm15[1],xmm18[1],xmm15[2],xmm18[2],xmm15[3],xmm18[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm18[0],xmm1[1],xmm18[1],xmm1[2],xmm18[2],xmm1[3],xmm18[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm15[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm0, %zmm15
+; AVX512DQ-BW-NEXT: vpmovqb %zmm15, %xmm15
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm15 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm3, %ymm2
+; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm4, %zmm18
+; AVX512DQ-BW-NEXT: vpermw %zmm18, %zmm17, %zmm8
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm8[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm5, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm6, %xmm8
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm18[0],xmm8[1],xmm18[1],xmm8[2],xmm18[2],xmm8[3],xmm18[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm5, %ymm18
-; AVX512DQ-BW-NEXT: vpmovqb %ymm18, %xmm0
-; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm0
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm15[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm6, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm9, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm10, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1,2],xmm2[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm19, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm8[0,1],xmm2[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm21
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm7, %xmm15
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm14, %xmm15
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm5, %ymm15
-; AVX512DQ-BW-NEXT: vpmovqb %ymm15, %xmm15
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm19, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm7, %xmm15
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm14, %xmm15
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm5, %ymm15
-; AVX512DQ-BW-NEXT: vpmovqb %ymm15, %xmm15
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm19, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm23
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm4, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm7, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm18[0],xmm11[0],xmm18[1],xmm11[1],xmm18[2],xmm11[2],xmm18[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm11
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm14, %xmm18
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm18[0],xmm2[1],xmm18[1],xmm2[2],xmm18[2],xmm2[3],xmm18[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm5, %ymm18
-; AVX512DQ-BW-NEXT: vpmovqb %ymm18, %xmm15
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm15
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3,4,5],ymm11[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm19, %zmm8
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm7, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm9, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm18[0],xmm1[1],xmm18[1],xmm1[2],xmm18[2],xmm1[3],xmm18[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm8[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm0, %zmm8
; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm24
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm4, %xmm15
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm7, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm17, %ymm15
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm14, %xmm18
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm11, %xmm8
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm10, %xmm2
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm3, %ymm8
+; AVX512DQ-BW-NEXT: vpmovqb %ymm8, %xmm8
+; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm8
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %zmm2, %zmm4, %zmm18
+; AVX512DQ-BW-NEXT: vpermw %zmm18, %zmm17, %zmm12
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm5, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm12, %xmm6, %xmm12
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm18[0],xmm12[1],xmm18[1],xmm12[2],xmm18[2],xmm12[3],xmm18[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm7, %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm9, %xmm2
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm18[0],xmm2[1],xmm18[1],xmm2[2],xmm18[2],xmm2[3],xmm18[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm5, %ymm18
-; AVX512DQ-BW-NEXT: vpmovqb %ymm18, %xmm11
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm16, %ymm11
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3,4,5],ymm15[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm10, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm12, %xmm8
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1,2],xmm12[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm0, %zmm12
+; AVX512DQ-BW-NEXT: vpmovqb %zmm12, %xmm12
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm12[0,1],xmm2[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm11, %xmm11
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm10, %xmm8
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1],xmm8[2],xmm11[2],xmm8[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm19, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm0, %ymm25
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm4, %xmm15
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm7, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm17, %ymm15
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm14, %xmm18
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm13, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm18[0],xmm8[1],xmm18[1],xmm8[2],xmm18[2],xmm8[3],xmm18[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm5, %ymm18
-; AVX512DQ-BW-NEXT: vpmovqb %ymm18, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm15[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm10, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm11, %xmm12, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm19, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm8[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm3, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm4, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm7, %xmm4
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm17, %ymm3
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm14, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm13, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm5, %ymm4
-; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
-; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm16, %ymm4
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm3[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm6, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm9, %xmm0
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm10, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm19, %zmm2
-; AVX512DQ-BW-NEXT: vpmovqb %zmm2, %xmm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm20, (%rsi)
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm21, (%rdx)
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm22, (%rcx)
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm23, (%r8)
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm24, (%r9)
-; AVX512DQ-BW-NEXT: vmovdqa64 %ymm25, (%r11)
-; AVX512DQ-BW-NEXT: vmovdqa %ymm8, (%r10)
+; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm3, %ymm3
+; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm3
+; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm16, %ymm3
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %zmm8, %zmm4, %zmm4
+; AVX512DQ-BW-NEXT: vpermw %zmm4, %zmm17, %zmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm4[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm5, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm6, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm7, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm9, %xmm6
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm0, %zmm0
+; AVX512DQ-BW-NEXT: vpmovqb %zmm0, %xmm0
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqa64 %ymm19, (%rsi)
+; AVX512DQ-BW-NEXT: vmovdqa64 %ymm20, (%rdx)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm13, (%rcx)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm14, (%r8)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm15, (%r9)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm1, (%r11)
+; AVX512DQ-BW-NEXT: vmovdqa %ymm2, (%r10)
; AVX512DQ-BW-NEXT: vmovdqa %ymm0, (%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
@@ -14882,509 +14790,408 @@ define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX512BW-LABEL: load_i8_stride8_vf64:
; AVX512BW: # %bb.0:
-; AVX512BW-NEXT: subq $392, %rsp # imm = 0x188
-; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm0
-; AVX512BW-NEXT: vpmovqb %zmm0, %xmm0
-; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm9
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512BW-NEXT: subq $264, %rsp # imm = 0x108
+; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm0
+; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm1
+; AVX512BW-NEXT: vpmovqb %zmm1, %xmm2
+; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm8
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT: vmovdqa 368(%rdi), %xmm10
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm10, %xmm1
-; AVX512BW-NEXT: vmovdqa 352(%rdi), %xmm11
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm11, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm6 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512BW-NEXT: vmovdqa 336(%rdi), %xmm12
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm12, %xmm2
-; AVX512BW-NEXT: vmovdqa 320(%rdi), %xmm13
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm13, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
-; AVX512BW-NEXT: vpmovqb %zmm9, %xmm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm2[0,1],xmm1[2,3]
-; AVX512BW-NEXT: vmovdqa 496(%rdi), %xmm14
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm14, %xmm1
-; AVX512BW-NEXT: vmovdqa64 480(%rdi), %xmm16
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm16, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vmovdqa64 464(%rdi), %xmm21
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm21, %xmm2
-; AVX512BW-NEXT: vmovdqa64 448(%rdi), %xmm22
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm22, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm25 = [2,11,2,11,2,11,2,11]
-; AVX512BW-NEXT: vpermt2d %ymm1, %ymm25, %ymm2
-; AVX512BW-NEXT: vmovdqa64 384(%rdi), %ymm17
-; AVX512BW-NEXT: vpmovqb %ymm17, %xmm1
-; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm30 = [0,9,0,9,0,9,0,9]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm2[6,7]
-; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = [0,1,10,11,0,1,10,11]
-; AVX512BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm4
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm10, %xmm1
+; AVX512BW-NEXT: vmovdqa 352(%rdi), %xmm12
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm12, %xmm3
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512BW-NEXT: vmovdqa 336(%rdi), %xmm14
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm14, %xmm4
+; AVX512BW-NEXT: vmovdqa64 320(%rdi), %xmm18
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm18, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpmovqb %zmm8, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vmovdqa 384(%rdi), %ymm9
+; AVX512BW-NEXT: vpmovqb %ymm9, %xmm3
+; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm29 = [0,9,0,9,0,9,0,9]
+; AVX512BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm3
+; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [18,2,27,11,18,2,27,11,18,2,27,11,18,2,27,11]
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm0[2,3,6,7,0,1,4,5]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm15, %zmm0
+; AVX512BW-NEXT: vpermw %zmm0, %zmm13, %zmm0
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7]
+; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [0,1,10,11,0,1,10,11]
+; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm4
; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vmovdqa64 416(%rdi), %xmm26
-; AVX512BW-NEXT: vmovdqa64 432(%rdi), %xmm27
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm28 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm31 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $8, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
+; AVX512BW-NEXT: vmovdqa64 416(%rdi), %xmm23
+; AVX512BW-NEXT: vmovdqa64 432(%rdi), %xmm24
; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $8, %ymm17, %ymm2
-; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm20 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $16, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $16, %ymm17, %ymm2
-; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $24, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $24, %ymm17, %ymm2
-; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $32, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm10, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $32, %ymm17, %ymm2
-; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $40, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm14, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm18, %xmm3
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
+; AVX512BW-NEXT: vpsrlq $8, %zmm8, %zmm2
+; AVX512BW-NEXT: vpmovqb %zmm2, %xmm2
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm2[0,1],xmm0[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm24, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm23, %xmm2
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $40, %ymm17, %ymm2
+; AVX512BW-NEXT: vpsrlq $8, %ymm9, %ymm2
; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512BW-NEXT: vmovdqu64 %zmm7, (%rsp) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm24 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm11, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm23 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm12, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm13, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $48, %zmm9, %zmm1
-; AVX512BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm1[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm16, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm21, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm22, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $48, %ymm17, %ymm3
-; AVX512BW-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm4
-; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm19 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm10, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm11, %xmm2
+; AVX512BW-NEXT: vpermt2d %ymm0, %ymm29, %ymm2
+; AVX512BW-NEXT: vpshufb %zmm4, %zmm15, %zmm0
+; AVX512BW-NEXT: vpermw %zmm0, %zmm13, %zmm0
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm3
+; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm10, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm15 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm12, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm13, %xmm3
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm20 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm20, %xmm14, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm20, %xmm18, %xmm3
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $56, %zmm9, %zmm2
+; AVX512BW-NEXT: vpsrlq $16, %zmm8, %zmm2
; AVX512BW-NEXT: vpmovqb %zmm2, %xmm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm0[2,3]
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm14, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm16, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm21, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm22, %xmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm24, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm23, %xmm3
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm29 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm27, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm26, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $56, %ymm17, %ymm3
+; AVX512BW-NEXT: vpsrlq $16, %ymm9, %ymm3
; AVX512BW-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm1
-; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512BW-NEXT: vmovdqa 240(%rdi), %xmm4
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm4, %xmm0
-; AVX512BW-NEXT: vmovdqa64 224(%rdi), %xmm18
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm18, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vmovdqa64 208(%rdi), %xmm22
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm22, %xmm2
-; AVX512BW-NEXT: vmovdqa64 192(%rdi), %xmm26
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm26, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm0
-; AVX512BW-NEXT: vpmovqb %zmm0, %xmm0
-; AVX512BW-NEXT: vmovdqa64 128(%rdi), %ymm17
-; AVX512BW-NEXT: vpmovqb %ymm17, %xmm3
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512BW-NEXT: vmovdqa64 112(%rdi), %xmm27
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm27, %xmm2
-; AVX512BW-NEXT: vmovdqa 96(%rdi), %xmm12
-; AVX512BW-NEXT: vpshufb %xmm5, %xmm12, %xmm3
+; AVX512BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm3
+; AVX512BW-NEXT: vpshufb %zmm20, %zmm15, %zmm2
+; AVX512BW-NEXT: vpermw %zmm2, %zmm13, %zmm2
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
+; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm10, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm12, %xmm3
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vmovdqa 80(%rdi), %xmm13
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm13, %xmm3
-; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm14
-; AVX512BW-NEXT: vpshufb %xmm6, %xmm14, %xmm5
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm21 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm21, %xmm14, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm21, %xmm18, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
-; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm16
-; AVX512BW-NEXT: vpmovqb %zmm16, %xmm3
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm2[0,1,2,3],ymm0[4,5,6,7]
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm4, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm28, %xmm18, %xmm2
-; AVX512BW-NEXT: vmovdqa64 %xmm28, %xmm9
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vmovdqa64 %xmm31, %xmm10
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm22, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm31, %xmm26, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm11
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm4, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm18, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm21 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm21, %xmm22, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm21, %xmm26, %xmm3
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm25, %ymm0
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm4, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm18, %xmm3
+; AVX512BW-NEXT: vpsrlq $24, %zmm8, %zmm3
+; AVX512BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm2[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm22 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm24, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm23, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $24, %ymm9, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %zmm21, %zmm15, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm10, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm27, %xmm14, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm27, %xmm18, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $32, %zmm8, %zmm4
+; AVX512BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm17 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm24, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm23, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $32, %ymm9, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %zmm27, %zmm15, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm10, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm31 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm31, %xmm14, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm31, %xmm18, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $40, %zmm8, %zmm4
+; AVX512BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm6 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm24, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm23, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $40, %ymm9, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %zmm31, %zmm15, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm30 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %xmm30, %xmm10, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm30, %xmm12, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm19 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm14, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm18, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $48, %zmm8, %zmm4
+; AVX512BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm11 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm28 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512BW-NEXT: vpshufb %xmm28, %xmm24, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm28, %xmm23, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $48, %ymm9, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %zmm19, %zmm15, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm11
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm26 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %xmm26, %xmm10, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm26, %xmm12, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm10 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm14, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm18, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $56, %zmm8, %zmm4
+; AVX512BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm25 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm24, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm23, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $56, %ymm9, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %zmm10, %zmm15, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm8
+; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm3
+; AVX512BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512BW-NEXT: vmovdqa64 128(%rdi), %ymm23
+; AVX512BW-NEXT: vpmovqb %ymm23, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm3
+; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm3[2,3,6,7,0,1,4,5]
+; AVX512BW-NEXT: vpshufb %zmm1, %zmm24, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vmovdqa 112(%rdi), %xmm12
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm12, %xmm4
+; AVX512BW-NEXT: vmovdqa 96(%rdi), %xmm14
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm14, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vmovdqa 80(%rdi), %xmm15
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm15, %xmm5
+; AVX512BW-NEXT: vmovdqa64 64(%rdi), %xmm16
+; AVX512BW-NEXT: vpshufb %xmm1, %xmm16, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3]
+; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm18
+; AVX512BW-NEXT: vpmovqb %zmm18, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm7 = ymm1[0,1,2,3],ymm3[4,5,6,7]
+; AVX512BW-NEXT: vmovdqa 160(%rdi), %xmm1
+; AVX512BW-NEXT: vmovdqa 176(%rdi), %xmm0
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm1, %xmm3
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $8, %ymm23, %ymm3
+; AVX512BW-NEXT: vpmovqb %ymm3, %xmm9
+; AVX512BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm9
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm1, %xmm3
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm22, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm5
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512BW-NEXT: vpermt2d %ymm3, %ymm25, %ymm7
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm4, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm18, %xmm5
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm22, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm6
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX512BW-NEXT: vpermt2d %ymm5, %ymm25, %ymm3
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm4, %xmm5
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm18, %xmm6
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm22, %xmm5
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm31
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm31[0],xmm5[0],xmm31[1],xmm5[1],xmm31[2],xmm5[2],xmm31[3],xmm5[3]
-; AVX512BW-NEXT: vpermt2d %ymm6, %ymm25, %ymm5
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm4, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm18, %xmm31
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm31 = xmm31[0],xmm6[0],xmm31[1],xmm6[1],xmm31[2],xmm6[2],xmm31[3],xmm6[3]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm22, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm26, %xmm28
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm28[0],xmm6[0],xmm28[1],xmm6[1],xmm28[2],xmm6[2],xmm28[3],xmm6[3]
-; AVX512BW-NEXT: vpermt2d %ymm31, %ymm25, %ymm6
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm4, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm18, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm18 = xmm18[0],xmm4[0],xmm18[1],xmm4[1],xmm18[2],xmm4[2],xmm18[3],xmm4[3]
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm22, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm26, %xmm22
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm22[0],xmm4[0],xmm22[1],xmm4[1],xmm22[2],xmm4[2],xmm22[3],xmm4[3]
-; AVX512BW-NEXT: vpermt2d %ymm18, %ymm25, %ymm4
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm1, %zmm25
-; AVX512BW-NEXT: vmovdqa64 160(%rdi), %xmm26
-; AVX512BW-NEXT: vmovdqa64 176(%rdi), %xmm28
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm28, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm18[0],xmm8[0],xmm18[1],xmm8[1],xmm18[2],xmm8[2],xmm18[3],xmm8[3]
-; AVX512BW-NEXT: vpsrlq $8, %ymm17, %ymm18
-; AVX512BW-NEXT: vpmovqb %ymm18, %xmm1
-; AVX512BW-NEXT: vpermt2d %ymm8, %ymm30, %ymm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm11[6,7]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm27, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm12, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpshufb %xmm10, %xmm13, %xmm11
-; AVX512BW-NEXT: vpshufb %xmm10, %xmm14, %xmm18
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm18[0],xmm11[0],xmm18[1],xmm11[1],xmm18[2],xmm11[2],xmm18[3],xmm11[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1,2],xmm8[3]
-; AVX512BW-NEXT: vpsrlq $8, %zmm16, %zmm11
-; AVX512BW-NEXT: vpmovqb %zmm11, %xmm11
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1],xmm8[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm18
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm28, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
-; AVX512BW-NEXT: vpsrlq $16, %ymm17, %ymm8
-; AVX512BW-NEXT: vpmovqb %ymm8, %xmm8
-; AVX512BW-NEXT: vpermt2d %ymm1, %ymm30, %ymm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],ymm0[6,7]
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm27, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm20, %xmm12, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
-; AVX512BW-NEXT: vpshufb %xmm21, %xmm13, %xmm8
-; AVX512BW-NEXT: vpshufb %xmm21, %xmm14, %xmm11
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3]
-; AVX512BW-NEXT: vpsrlq $16, %zmm16, %zmm8
-; AVX512BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm22
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm28, %xmm0
-; AVX512BW-NEXT: vpshufb %xmm1, %xmm26, %xmm1
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512BW-NEXT: vpsrlq $24, %ymm17, %ymm1
-; AVX512BW-NEXT: vpmovqb %ymm1, %xmm0
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm0
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm28, %xmm1
-; AVX512BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512BW-NEXT: vpsrlq $32, %ymm17, %ymm1
-; AVX512BW-NEXT: vpmovqb %ymm1, %xmm1
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm1
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm28, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm26, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpsrlq $40, %ymm17, %ymm8
-; AVX512BW-NEXT: vpmovqb %ymm8, %xmm11
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm11
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm28, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm8, %xmm26, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512BW-NEXT: vpsrlq $48, %ymm17, %ymm8
-; AVX512BW-NEXT: vpmovqb %ymm8, %xmm8
-; AVX512BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm8
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm28, %xmm2
-; AVX512BW-NEXT: vpshufb %xmm29, %xmm26, %xmm21
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm21 = xmm21[0],xmm2[0],xmm21[1],xmm2[1],xmm21[2],xmm2[2],xmm21[3],xmm2[3]
-; AVX512BW-NEXT: vpsrlq $56, %ymm17, %ymm2
+; AVX512BW-NEXT: vpsrlq $16, %ymm23, %ymm2
; AVX512BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512BW-NEXT: vpermt2d %ymm21, %ymm30, %ymm2
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm27, %xmm7
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm12, %xmm17
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm17[0],xmm7[0],xmm17[1],xmm7[1],xmm17[2],xmm7[2],xmm17[3],xmm7[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm13, %xmm17
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm14, %xmm21
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm21[0],xmm17[0],xmm21[1],xmm17[1],xmm21[2],xmm17[2],xmm21[3],xmm17[3]
+; AVX512BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm2
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm0, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm1, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $24, %ymm23, %ymm3
+; AVX512BW-NEXT: vpmovqb %ymm3, %xmm3
+; AVX512BW-NEXT: vpermt2d %ymm4, %ymm29, %ymm3
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm0, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm17, %xmm1, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpsrlq $32, %ymm23, %ymm4
+; AVX512BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512BW-NEXT: vpermt2d %ymm5, %ymm29, %ymm4
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm0, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm6, %xmm1, %xmm6
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512BW-NEXT: vpsrlq $40, %ymm23, %ymm6
+; AVX512BW-NEXT: vpmovqb %ymm6, %xmm6
+; AVX512BW-NEXT: vpermt2d %ymm5, %ymm29, %ymm6
+; AVX512BW-NEXT: vpshufb %xmm28, %xmm0, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm28, %xmm1, %xmm17
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm5[0],xmm17[1],xmm5[1],xmm17[2],xmm5[2],xmm17[3],xmm5[3]
+; AVX512BW-NEXT: vpsrlq $48, %ymm23, %ymm5
+; AVX512BW-NEXT: vpmovqb %ymm5, %xmm5
+; AVX512BW-NEXT: vpermt2d %ymm17, %ymm29, %ymm5
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm0, %xmm0
+; AVX512BW-NEXT: vpshufb %xmm25, %xmm1, %xmm1
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512BW-NEXT: vpsrlq $56, %ymm23, %ymm0
+; AVX512BW-NEXT: vpmovqb %ymm0, %xmm0
+; AVX512BW-NEXT: vpermt2d %ymm1, %ymm29, %ymm0
+; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm1, %zmm17
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} zmm22 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %zmm22, %zmm24, %zmm1
+; AVX512BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5],ymm1[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512BW-NEXT: vpshufb %xmm9, %xmm12, %xmm7
+; AVX512BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm15, %xmm9
+; AVX512BW-NEXT: vpshufb %xmm22, %xmm16, %xmm22
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm22[0],xmm9[0],xmm22[1],xmm9[1],xmm22[2],xmm9[2],xmm22[3],xmm9[3]
; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm9[0,1,2],xmm7[3]
-; AVX512BW-NEXT: vpsrlq $24, %zmm16, %zmm9
+; AVX512BW-NEXT: vpsrlq $8, %zmm18, %zmm9
; AVX512BW-NEXT: vpmovqb %zmm9, %xmm9
; AVX512BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm9[0,1],xmm7[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm27, %xmm3
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm12, %xmm7
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm13, %xmm7
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm7[0,1,2],xmm3[3]
-; AVX512BW-NEXT: vpsrlq $32, %zmm16, %zmm7
-; AVX512BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm7[0,1],xmm3[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm11[0,1,2,3,4,5],ymm5[6,7]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm27, %xmm5
-; AVX512BW-NEXT: vpshufb %xmm7, %xmm12, %xmm7
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm13, %xmm7
-; AVX512BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm22
+; AVX512BW-NEXT: vpshufb %zmm20, %zmm24, %zmm1
+; AVX512BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm12, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm7, %xmm14, %xmm7
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
+; AVX512BW-NEXT: vpshufb %xmm20, %xmm15, %xmm7
+; AVX512BW-NEXT: vpshufb %xmm20, %xmm16, %xmm9
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm5 = xmm7[0,1,2],xmm5[3]
-; AVX512BW-NEXT: vpsrlq $40, %zmm16, %zmm7
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm7[0,1,2],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $16, %zmm18, %zmm7
; AVX512BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm5 = xmm7[0,1],xmm5[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm5 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm6[6,7]
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm27, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm24, %xmm12, %xmm7
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm13, %xmm7
-; AVX512BW-NEXT: vpshufb %xmm23, %xmm14, %xmm8
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
-; AVX512BW-NEXT: vpsrlq $48, %zmm16, %zmm7
-; AVX512BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm27, %xmm4
-; AVX512BW-NEXT: vpshufb %xmm19, %xmm12, %xmm6
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm20
+; AVX512BW-NEXT: vpshufb %zmm21, %zmm24, %zmm1
+; AVX512BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm12, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm14, %xmm3
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512BW-NEXT: vpshufb %xmm21, %xmm15, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm21, %xmm16, %xmm7
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $24, %zmm18, %zmm3
+; AVX512BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm21
+; AVX512BW-NEXT: vpshufb %zmm27, %zmm24, %zmm1
+; AVX512BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm12, %xmm2
+; AVX512BW-NEXT: vpshufb %xmm3, %xmm14, %xmm3
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512BW-NEXT: vpshufb %xmm27, %xmm15, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm27, %xmm16, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
+; AVX512BW-NEXT: vpsrlq $32, %zmm18, %zmm3
+; AVX512BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
+; AVX512BW-NEXT: vpshufb %zmm31, %zmm24, %zmm2
+; AVX512BW-NEXT: vpermw %zmm2, %zmm13, %zmm2
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm2[6,7]
+; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm12, %xmm3
+; AVX512BW-NEXT: vpshufb %xmm4, %xmm14, %xmm4
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512BW-NEXT: vpshufb %xmm31, %xmm15, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm31, %xmm16, %xmm6
; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm13, %xmm6
-; AVX512BW-NEXT: vpshufb %xmm15, %xmm14, %xmm7
-; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3]
-; AVX512BW-NEXT: vpsrlq $56, %zmm16, %zmm6
-; AVX512BW-NEXT: vpmovqb %zmm6, %xmm6
-; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
-; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]
-; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
-; AVX512BW-NEXT: vmovdqa64 %zmm25, (%rsi)
-; AVX512BW-NEXT: vmovdqa64 %zmm18, (%rdx)
-; AVX512BW-NEXT: vmovdqa64 %zmm22, (%rcx)
-; AVX512BW-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512BW-NEXT: vpsrlq $40, %zmm18, %zmm4
+; AVX512BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
+; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
+; AVX512BW-NEXT: vpshufb %zmm19, %zmm24, %zmm3
+; AVX512BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]
+; AVX512BW-NEXT: vpshufb %xmm30, %xmm12, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm30, %xmm14, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm15, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm19, %xmm16, %xmm6
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512BW-NEXT: vpsrlq $48, %zmm18, %zmm5
+; AVX512BW-NEXT: vpmovqb %zmm5, %xmm5
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm3
+; AVX512BW-NEXT: vpshufb %zmm10, %zmm24, %zmm4
+; AVX512BW-NEXT: vpermw %zmm4, %zmm13, %zmm4
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]
+; AVX512BW-NEXT: vpshufb %xmm26, %xmm12, %xmm4
+; AVX512BW-NEXT: vpshufb %xmm26, %xmm14, %xmm5
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm15, %xmm5
+; AVX512BW-NEXT: vpshufb %xmm10, %xmm16, %xmm6
+; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512BW-NEXT: vpsrlq $56, %zmm18, %zmm5
+; AVX512BW-NEXT: vpmovqb %zmm5, %xmm5
+; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
+; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
+; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
+; AVX512BW-NEXT: vmovdqa64 %zmm17, (%rsi)
+; AVX512BW-NEXT: vmovdqa64 %zmm22, (%rdx)
+; AVX512BW-NEXT: vmovdqa64 %zmm20, (%rcx)
+; AVX512BW-NEXT: vmovdqa64 %zmm21, (%r8)
; AVX512BW-NEXT: vmovdqa64 %zmm1, (%r9)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512BW-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512BW-NEXT: addq $264, %rsp # imm = 0x108
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
@@ -15737,509 +15544,408 @@ define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf64:
; AVX512DQ-BW: # %bb.0:
-; AVX512DQ-BW-NEXT: subq $392, %rsp # imm = 0x188
-; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm0
-; AVX512DQ-BW-NEXT: vpmovqb %zmm0, %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm9
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512DQ-BW-NEXT: subq $264, %rsp # imm = 0x108
+; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm0
+; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm1
+; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm2
+; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm8
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT: vmovdqa 368(%rdi), %xmm10
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm10, %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa 352(%rdi), %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm11, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm6 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
-; AVX512DQ-BW-NEXT: vmovdqa 336(%rdi), %xmm12
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm12, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa 320(%rdi), %xmm13
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm13, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
-; AVX512DQ-BW-NEXT: vpmovqb %zmm9, %xmm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm2[0,1],xmm1[2,3]
-; AVX512DQ-BW-NEXT: vmovdqa 496(%rdi), %xmm14
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm14, %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa64 480(%rdi), %xmm16
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm16, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vmovdqa64 464(%rdi), %xmm21
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm21, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %xmm22
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm22, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm25 = [2,11,2,11,2,11,2,11]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm25, %ymm2
-; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %ymm17
-; AVX512DQ-BW-NEXT: vpmovqb %ymm17, %xmm1
-; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm30 = [0,9,0,9,0,9,0,9]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm2[6,7]
-; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm18 = [0,1,10,11,0,1,10,11]
-; AVX512DQ-BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm10, %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa 352(%rdi), %xmm12
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm12, %xmm3
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
+; AVX512DQ-BW-NEXT: vmovdqa 336(%rdi), %xmm14
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %xmm18
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm18, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vmovdqa 384(%rdi), %ymm9
+; AVX512DQ-BW-NEXT: vpmovqb %ymm9, %xmm3
+; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm29 = [0,9,0,9,0,9,0,9]
+; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm3
+; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [18,2,27,11,18,2,27,11,18,2,27,11,18,2,27,11]
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm15 = zmm0[2,3,6,7,0,1,4,5]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm15, %zmm0
+; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm13, %zmm0
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm0[6,7]
+; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [0,1,10,11,0,1,10,11]
+; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm4
; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vmovdqa64 416(%rdi), %xmm26
-; AVX512DQ-BW-NEXT: vmovdqa64 432(%rdi), %xmm27
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm28 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm31 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
+; AVX512DQ-BW-NEXT: vmovdqa64 416(%rdi), %xmm23
+; AVX512DQ-BW-NEXT: vmovdqa64 432(%rdi), %xmm24
; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm17, %ymm2
-; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm20 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm17, %ymm2
-; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm17, %ymm2
-; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm10, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm17, %ymm2
-; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm14, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm18, %xmm3
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
+; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm8, %zmm2
+; AVX512DQ-BW-NEXT: vpmovqb %zmm2, %xmm2
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm2[0,1],xmm0[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm24, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm23, %xmm2
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm17, %ymm2
+; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm9, %ymm2
; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm7
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, (%rsp) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm24 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm11, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm23 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm12, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm13, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm9, %zmm1
-; AVX512DQ-BW-NEXT: vpmovqb %zmm1, %xmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm1[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm16, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm21, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm17, %ymm3
-; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm4
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm19 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm10, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm11, %xmm2
+; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm29, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %zmm4, %zmm15, %zmm0
+; AVX512DQ-BW-NEXT: vpermw %zmm0, %zmm13, %zmm0
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm3
+; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm10, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm12, %xmm2
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm15 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm12, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm13, %xmm3
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm20 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm14, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm18, %xmm3
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm9, %zmm2
+; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm8, %zmm2
; AVX512DQ-BW-NEXT: vpmovqb %zmm2, %xmm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm0[2,3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm14, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm16, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm21, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm22, %xmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm24, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm23, %xmm3
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm29 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm27, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm26, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm17, %ymm3
+; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm9, %ymm3
; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm3
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm18, %zmm1
-; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-BW-NEXT: vmovdqa 240(%rdi), %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm4, %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa64 224(%rdi), %xmm18
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm18, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vmovdqa64 208(%rdi), %xmm22
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %xmm26
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm26, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm2
-; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm0
-; AVX512DQ-BW-NEXT: vpmovqb %zmm0, %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %ymm17
-; AVX512DQ-BW-NEXT: vpmovqb %ymm17, %xmm3
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm30, %ymm3
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]
-; AVX512DQ-BW-NEXT: vmovdqa64 112(%rdi), %xmm27
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm27, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa 96(%rdi), %xmm12
-; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm12, %xmm3
+; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm3
+; AVX512DQ-BW-NEXT: vpshufb %zmm20, %zmm15, %zmm2
+; AVX512DQ-BW-NEXT: vpermw %zmm2, %zmm13, %zmm2
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm10, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm12, %xmm3
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vmovdqa 80(%rdi), %xmm13
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm13, %xmm3
-; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm14
-; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm14, %xmm5
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm21 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm14, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm18, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm16
-; AVX512DQ-BW-NEXT: vpmovqb %zmm16, %xmm3
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm2[0,1,2,3],ymm0[4,5,6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm4, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm18, %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa64 %xmm28, %xmm9
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vmovdqa64 %xmm31, %xmm10
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm26, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm0, %ymm25, %ymm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm4, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm18, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm21 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm22, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm26, %xmm3
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm25, %ymm0
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm4, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm18, %xmm3
+; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm8, %zmm3
+; AVX512DQ-BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm2[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm22 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm24, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm23, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm9, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %zmm21, %zmm15, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm10, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm18, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm8, %zmm4
+; AVX512DQ-BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm17 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm24, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm23, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm9, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %zmm27, %zmm15, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm10, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm31 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm18, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm8, %zmm4
+; AVX512DQ-BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm6 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm24, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm23, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm9, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %zmm31, %zmm15, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm30 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %xmm30, %xmm10, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm30, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm19 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm18, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm8, %zmm4
+; AVX512DQ-BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm11 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm28 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
+; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm24, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm23, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm9, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %zmm19, %zmm15, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm11
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm26 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm10, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm10 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm18, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm8, %zmm4
+; AVX512DQ-BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm25 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm24, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm23, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm9, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %zmm10, %zmm15, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm8
+; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm3
+; AVX512DQ-BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %ymm23
+; AVX512DQ-BW-NEXT: vpmovqb %ymm23, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm3
+; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm24 = zmm3[2,3,6,7,0,1,4,5]
+; AVX512DQ-BW-NEXT: vpshufb %zmm1, %zmm24, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vmovdqa 112(%rdi), %xmm12
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vmovdqa 96(%rdi), %xmm14
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm14, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vmovdqa 80(%rdi), %xmm15
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm15, %xmm5
+; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %xmm16
+; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm16, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3]
+; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm18
+; AVX512DQ-BW-NEXT: vpmovqb %zmm18, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm7 = ymm1[0,1,2,3],ymm3[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqa 160(%rdi), %xmm1
+; AVX512DQ-BW-NEXT: vmovdqa 176(%rdi), %xmm0
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm1, %xmm3
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm23, %ymm3
+; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm9
+; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm29, %ymm9
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm0, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm1, %xmm3
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm22, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm5
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm25, %ymm7
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm4, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm18, %xmm5
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm22, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm6
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm5, %ymm25, %ymm3
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm4, %xmm5
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm18, %xmm6
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm22, %xmm5
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm31
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm31[0],xmm5[0],xmm31[1],xmm5[1],xmm31[2],xmm5[2],xmm31[3],xmm5[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm6, %ymm25, %ymm5
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm4, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm18, %xmm31
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm31 = xmm31[0],xmm6[0],xmm31[1],xmm6[1],xmm31[2],xmm6[2],xmm31[3],xmm6[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm22, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm26, %xmm28
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm28[0],xmm6[0],xmm28[1],xmm6[1],xmm28[2],xmm6[2],xmm28[3],xmm6[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm31, %ymm25, %ymm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm4, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm18, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm18 = xmm18[0],xmm4[0],xmm18[1],xmm4[1],xmm18[2],xmm4[2],xmm18[3],xmm4[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm22, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm26, %xmm22
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm22[0],xmm4[0],xmm22[1],xmm4[1],xmm22[2],xmm4[2],xmm22[3],xmm4[3]
-; AVX512DQ-BW-NEXT: vpermt2d %ymm18, %ymm25, %ymm4
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm1, %zmm25
-; AVX512DQ-BW-NEXT: vmovdqa64 160(%rdi), %xmm26
-; AVX512DQ-BW-NEXT: vmovdqa64 176(%rdi), %xmm28
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm28, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm18[0],xmm8[0],xmm18[1],xmm8[1],xmm18[2],xmm8[2],xmm18[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %ymm17, %ymm18
-; AVX512DQ-BW-NEXT: vpmovqb %ymm18, %xmm1
-; AVX512DQ-BW-NEXT: vpermt2d %ymm8, %ymm30, %ymm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm11[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm27, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm12, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm13, %xmm11
-; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm14, %xmm18
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm18[0],xmm11[0],xmm18[1],xmm11[1],xmm18[2],xmm11[2],xmm18[3],xmm11[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1,2],xmm8[3]
-; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm16, %zmm11
-; AVX512DQ-BW-NEXT: vpmovqb %zmm11, %xmm11
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm8 = xmm11[0,1],xmm8[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm18
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm28, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm17, %ymm8
-; AVX512DQ-BW-NEXT: vpmovqb %ymm8, %xmm8
-; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm30, %ymm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],ymm0[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm27, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm12, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm13, %xmm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm14, %xmm11
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1,2],xmm1[3]
-; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm16, %zmm8
-; AVX512DQ-BW-NEXT: vpmovqb %zmm8, %xmm8
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm22
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm28, %xmm0
-; AVX512DQ-BW-NEXT: vpshufb %xmm1, %xmm26, %xmm1
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm17, %ymm1
-; AVX512DQ-BW-NEXT: vpmovqb %ymm1, %xmm0
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm0
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm28, %xmm1
-; AVX512DQ-BW-NEXT: vpshufb %xmm2, %xmm26, %xmm2
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm17, %ymm1
-; AVX512DQ-BW-NEXT: vpmovqb %ymm1, %xmm1
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm1
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm28, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm26, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm17, %ymm8
-; AVX512DQ-BW-NEXT: vpmovqb %ymm8, %xmm11
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm11
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm28, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm8, %xmm26, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm17, %ymm8
-; AVX512DQ-BW-NEXT: vpmovqb %ymm8, %xmm8
-; AVX512DQ-BW-NEXT: vpermt2d %ymm2, %ymm30, %ymm8
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm28, %xmm2
-; AVX512DQ-BW-NEXT: vpshufb %xmm29, %xmm26, %xmm21
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm21 = xmm21[0],xmm2[0],xmm21[1],xmm2[1],xmm21[2],xmm2[2],xmm21[3],xmm2[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm17, %ymm2
+; AVX512DQ-BW-NEXT: vpsrlq $16, %ymm23, %ymm2
; AVX512DQ-BW-NEXT: vpmovqb %ymm2, %xmm2
-; AVX512DQ-BW-NEXT: vpermt2d %ymm21, %ymm30, %ymm2
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm27, %xmm7
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm12, %xmm17
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm17[0],xmm7[0],xmm17[1],xmm7[1],xmm17[2],xmm7[2],xmm17[3],xmm7[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm13, %xmm17
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm14, %xmm21
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm21[0],xmm17[0],xmm21[1],xmm17[1],xmm21[2],xmm17[2],xmm21[3],xmm17[3]
+; AVX512DQ-BW-NEXT: vpermt2d %ymm3, %ymm29, %ymm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm0, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm1, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $24, %ymm23, %ymm3
+; AVX512DQ-BW-NEXT: vpmovqb %ymm3, %xmm3
+; AVX512DQ-BW-NEXT: vpermt2d %ymm4, %ymm29, %ymm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm0, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm17, %xmm1, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %ymm23, %ymm4
+; AVX512DQ-BW-NEXT: vpmovqb %ymm4, %xmm4
+; AVX512DQ-BW-NEXT: vpermt2d %ymm5, %ymm29, %ymm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm0, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm6, %xmm1, %xmm6
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %ymm23, %ymm6
+; AVX512DQ-BW-NEXT: vpmovqb %ymm6, %xmm6
+; AVX512DQ-BW-NEXT: vpermt2d %ymm5, %ymm29, %ymm6
+; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm0, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm28, %xmm1, %xmm17
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm5[0],xmm17[1],xmm5[1],xmm17[2],xmm5[2],xmm17[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %ymm23, %ymm5
+; AVX512DQ-BW-NEXT: vpmovqb %ymm5, %xmm5
+; AVX512DQ-BW-NEXT: vpermt2d %ymm17, %ymm29, %ymm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm0, %xmm0
+; AVX512DQ-BW-NEXT: vpshufb %xmm25, %xmm1, %xmm1
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX512DQ-BW-NEXT: vpsrlq $56, %ymm23, %ymm0
+; AVX512DQ-BW-NEXT: vpmovqb %ymm0, %xmm0
+; AVX512DQ-BW-NEXT: vpermt2d %ymm1, %ymm29, %ymm0
+; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm1, %zmm17
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} zmm22 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %zmm22, %zmm24, %zmm1
+; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5],ymm1[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
+; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm12, %xmm7
+; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm15, %xmm9
+; AVX512DQ-BW-NEXT: vpshufb %xmm22, %xmm16, %xmm22
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm22[0],xmm9[0],xmm22[1],xmm9[1],xmm22[2],xmm9[2],xmm22[3],xmm9[3]
; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm9[0,1,2],xmm7[3]
-; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm16, %zmm9
+; AVX512DQ-BW-NEXT: vpsrlq $8, %zmm18, %zmm9
; AVX512DQ-BW-NEXT: vpmovqb %zmm9, %xmm9
; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm7 = xmm9[0,1],xmm7[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm27, %xmm3
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm12, %xmm7
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm13, %xmm7
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm7[0,1,2],xmm3[3]
-; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm16, %zmm7
-; AVX512DQ-BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm7[0,1],xmm3[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm11[0,1,2,3,4,5],ymm5[6,7]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm27, %xmm5
-; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm12, %xmm7
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm9 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm13, %xmm7
-; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm14, %xmm9
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm7, %zmm22
+; AVX512DQ-BW-NEXT: vpshufb %zmm20, %zmm24, %zmm1
+; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm12, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm14, %xmm7
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm15, %xmm7
+; AVX512DQ-BW-NEXT: vpshufb %xmm20, %xmm16, %xmm9
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm5 = xmm7[0,1,2],xmm5[3]
-; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm16, %zmm7
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm7[0,1,2],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $16, %zmm18, %zmm7
; AVX512DQ-BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm5 = xmm7[0,1],xmm5[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 (%rsp), %zmm5 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm6[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm27, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm24, %xmm12, %xmm7
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm13, %xmm7
-; AVX512DQ-BW-NEXT: vpshufb %xmm23, %xmm14, %xmm8
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
-; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm16, %zmm7
-; AVX512DQ-BW-NEXT: vpmovqb %zmm7, %xmm7
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm27, %xmm4
-; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm12, %xmm6
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm20
+; AVX512DQ-BW-NEXT: vpshufb %zmm21, %zmm24, %zmm1
+; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm12, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm14, %xmm3
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm15, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm21, %xmm16, %xmm7
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $24, %zmm18, %zmm3
+; AVX512DQ-BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu64 (%rsp), %zmm2 # 64-byte Reload
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm21
+; AVX512DQ-BW-NEXT: vpshufb %zmm27, %zmm24, %zmm1
+; AVX512DQ-BW-NEXT: vpermw %zmm1, %zmm13, %zmm1
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm12, %xmm2
+; AVX512DQ-BW-NEXT: vpshufb %xmm3, %xmm14, %xmm3
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm15, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm27, %xmm16, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
+; AVX512DQ-BW-NEXT: vpsrlq $32, %zmm18, %zmm3
+; AVX512DQ-BW-NEXT: vpmovqb %zmm3, %xmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
+; AVX512DQ-BW-NEXT: vpshufb %zmm31, %zmm24, %zmm2
+; AVX512DQ-BW-NEXT: vpermw %zmm2, %zmm13, %zmm2
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm2[6,7]
+; AVX512DQ-BW-NEXT: vpbroadcastw {{.*#+}} xmm4 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm12, %xmm3
+; AVX512DQ-BW-NEXT: vpshufb %xmm4, %xmm14, %xmm4
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm15, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm31, %xmm16, %xmm6
; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm13, %xmm6
-; AVX512DQ-BW-NEXT: vpshufb %xmm15, %xmm14, %xmm7
-; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3]
-; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm16, %zmm6
-; AVX512DQ-BW-NEXT: vpmovqb %zmm6, %xmm6
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
-; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]
-; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, (%rsi)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, (%rdx)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, (%rcx)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
+; AVX512DQ-BW-NEXT: vpsrlq $40, %zmm18, %zmm4
+; AVX512DQ-BW-NEXT: vpmovqb %zmm4, %xmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
+; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
+; AVX512DQ-BW-NEXT: vpshufb %zmm19, %zmm24, %zmm3
+; AVX512DQ-BW-NEXT: vpermw %zmm3, %zmm13, %zmm3
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm30, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm30, %xmm14, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm15, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm19, %xmm16, %xmm6
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512DQ-BW-NEXT: vpsrlq $48, %zmm18, %zmm5
+; AVX512DQ-BW-NEXT: vpmovqb %zmm5, %xmm5
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm3
+; AVX512DQ-BW-NEXT: vpshufb %zmm10, %zmm24, %zmm4
+; AVX512DQ-BW-NEXT: vpermw %zmm4, %zmm13, %zmm4
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]
+; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm12, %xmm4
+; AVX512DQ-BW-NEXT: vpshufb %xmm26, %xmm14, %xmm5
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm15, %xmm5
+; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm16, %xmm6
+; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
+; AVX512DQ-BW-NEXT: vpsrlq $56, %zmm18, %zmm5
+; AVX512DQ-BW-NEXT: vpmovqb %zmm5, %xmm5
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
+; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
+; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, (%rsi)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, (%rdx)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, (%rcx)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, (%r8)
; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, (%r9)
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%rax)
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-BW-NEXT: addq $392, %rsp # imm = 0x188
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rax)
+; AVX512DQ-BW-NEXT: addq $264, %rsp # imm = 0x108
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll
index fc4377a08d560..904d07cd3d66f 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
@@ -916,121 +916,109 @@ define void @store_i16_stride4_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-LABEL: store_i16_stride4_vf16:
; AVX512: # %bb.0:
-; AVX512-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512-NEXT: vmovdqa 16(%rcx), %xmm1
-; AVX512-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-NEXT: vmovdqa 16(%rsi), %xmm6
-; AVX512-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512-NEXT: vmovdqa 16(%rdi), %xmm8
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,8,1,9,2,10,3,11]
-; AVX512-NEXT: vpermt2d %ymm4, %ymm10, %ymm9
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512-NEXT: vpermt2d %ymm1, %ymm10, %ymm3
-; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm1
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512-NEXT: vpermt2d %ymm3, %ymm10, %ymm4
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512-NEXT: vpermt2d %ymm0, %ymm10, %ymm2
-; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm0
-; AVX512-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,8,1,9,2,10,3,11]
+; AVX512-NEXT: vpermt2d %ymm4, %ymm8, %ymm7
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512-NEXT: vpermt2d %ymm2, %ymm8, %ymm3
+; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm2
+; AVX512-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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]
+; AVX512-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,4,1,5,2,6,3,7,16,20,17,21,18,22,19,23]
+; AVX512-NEXT: vpermi2d %zmm3, %zmm0, %zmm1
; AVX512-NEXT: vmovdqa64 %zmm1, 64(%r8)
+; AVX512-NEXT: vmovdqa64 %zmm2, (%r8)
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
; AVX512-FCP-LABEL: store_i16_stride4_vf16:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512-FCP-NEXT: vmovdqa 16(%rcx), %xmm1
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512-FCP-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512-FCP-NEXT: vmovdqa 16(%rsi), %xmm6
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512-FCP-NEXT: vmovdqa 16(%rdi), %xmm8
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,8,1,9,2,10,3,11]
-; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm10, %ymm9
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm10, %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm1
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm10, %ymm4
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512-FCP-NEXT: vpermt2d %ymm0, %ymm10, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm0
-; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,8,1,9,2,10,3,11]
+; AVX512-FCP-NEXT: vpermt2d %ymm4, %ymm8, %ymm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512-FCP-NEXT: vpermt2d %ymm2, %ymm8, %ymm3
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm2
+; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,4,1,5,2,6,3,7,16,20,17,21,18,22,19,23]
+; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm1
; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 64(%r8)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%r8)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
; AVX512DQ-LABEL: store_i16_stride4_vf16:
; AVX512DQ: # %bb.0:
-; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512DQ-NEXT: vmovdqa 16(%rcx), %xmm1
-; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512DQ-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-NEXT: vmovdqa 16(%rsi), %xmm6
-; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512DQ-NEXT: vmovdqa 16(%rdi), %xmm8
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,8,1,9,2,10,3,11]
-; AVX512DQ-NEXT: vpermt2d %ymm4, %ymm10, %ymm9
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512DQ-NEXT: vpermt2d %ymm1, %ymm10, %ymm3
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm1
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-NEXT: vpermt2d %ymm3, %ymm10, %ymm4
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-NEXT: vpermt2d %ymm0, %ymm10, %ymm2
-; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm0
-; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,8,1,9,2,10,3,11]
+; AVX512DQ-NEXT: vpermt2d %ymm4, %ymm8, %ymm7
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-NEXT: vpermt2d %ymm2, %ymm8, %ymm3
+; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm2
+; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512DQ-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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-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-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,4,1,5,2,6,3,7,16,20,17,21,18,22,19,23]
+; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm0, %zmm1
; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%r8)
+; AVX512DQ-NEXT: vmovdqa64 %zmm2, (%r8)
; AVX512DQ-NEXT: vzeroupper
; AVX512DQ-NEXT: retq
;
; AVX512DQ-FCP-LABEL: store_i16_stride4_vf16:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa 16(%rcx), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdx), %xmm3
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa 16(%rsi), %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdi), %xmm8
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm10 = [0,8,1,9,2,10,3,11]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm10, %ymm9
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm10, %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm3, %zmm1
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm10, %ymm4
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpermt2d %ymm0, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%r8)
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm6
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [0,8,1,9,2,10,3,11]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm4, %ymm8, %ymm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; AVX512DQ-FCP-NEXT: vpermt2d %ymm2, %ymm8, %ymm3
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm3, %zmm2
+; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} ymm3 = 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-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,4,1,5,2,6,3,7,16,20,17,21,18,22,19,23]
+; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 64(%r8)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%r8)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll
index a4299ffe7a736..05bd6199dcd3d 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
@@ -2955,126 +2955,128 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-FCP-LABEL: store_i16_stride7_vf16:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm2
-; AVX512-FCP-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
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm16
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm11
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm12
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,14,15],zero,zero,ymm4[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm4[u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17],zero,zero,zero,zero
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm18
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [2,0,3,2,0,10,10,11]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm9
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[1,1,1,1,5,5,5,5]
-; AVX512-FCP-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]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm19
-; AVX512-FCP-NEXT: vprold $16, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,0,2,1,4,4,6,5]
-; AVX512-FCP-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]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,2,3,3,10,0,11,10]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm13
-; AVX512-FCP-NEXT: vprold $16, %xmm12, %xmm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2],xmm3[3,4],xmm0[5],xmm3[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,0,1,0,0,8,0,9]
-; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm15
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm14[1],xmm8[2,3],xmm14[4],xmm8[5,6],xmm14[7]
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm14
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm11
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm8
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm12
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm10
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10, %ymm10
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [0,0,0,1,8,9,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm14, %zmm8, %zmm20
-; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm10
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512-FCP-NEXT: 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]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT: vpermd %zmm8, %zmm6, %zmm6
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [6,0,0,0,7,0,0,7]
-; AVX512-FCP-NEXT: vpermd %ymm8, %ymm7, %ymm7
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm12[0,0,1,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,2,3]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rcx
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm2))
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm8, %zmm2
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm2 & mem)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm0))
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm19[2,2,2,3,6,6,6,7]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm9 ^ (mem & (zmm0 ^ zmm9))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm13))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm0))
-; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm11[0,0,1,1,4,4,5,5]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm15 ^ (mem & (zmm0 ^ zmm15))
-; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm2
-; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm20))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm5))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm7 ^ (mem & (ymm1 ^ ymm7))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm4))
-; AVX512-FCP-NEXT: vmovdqa %ymm1, 192(%rcx)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rcx)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rcx)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 64(%rcx)
+; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm1
+; AVX512-FCP-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
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
+; AVX512-FCP-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]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm10, %ymm7
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm12, %zmm14
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm6))
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[16,17],zero,zero,zero,zero
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
+; AVX512-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm12
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm13
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vmovdqa64 %xmm15, %xmm17
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm16
+; AVX512-FCP-NEXT: vpbroadcastd 8(%r10), %ymm6
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm6
+; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm15 = 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]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm16 & mem)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm14))
+; AVX512-FCP-NEXT: vprold $16, %xmm9, %xmm14
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm15 = xmm8[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0,1],xmm14[2],xmm15[3,4],xmm14[5],xmm15[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,1,0,0,8,0,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm14, %zmm15, %zmm16
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm16 ^ (mem & (zmm9 ^ zmm16))
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm8
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,16,17,18,19,24,25,26,27,30,31,28,29,30,31,28,29]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [0,0,0,1,2,3,3,0]
+; AVX512-FCP-NEXT: vpermq %zmm8, %zmm10, %zmm10
+; AVX512-FCP-NEXT: vpbroadcastd (%r10), %ymm8
+; AVX512-FCP-NEXT: vpbroadcastd 4(%r10), %ymm11
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm10))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm9))
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,0,3,2,0,10,10,11]
+; AVX512-FCP-NEXT: vpermi2q %zmm9, %zmm10, %zmm11
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512-FCP-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]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm10 = 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]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm11 ^ (mem & (zmm9 ^ zmm11))
+; AVX512-FCP-NEXT: vprold $16, %ymm1, %ymm10
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm12 = [2,2,3,3,10,0,11,10]
+; AVX512-FCP-NEXT: vpermi2q %zmm10, %zmm11, %zmm12
+; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vpermd %zmm7, %zmm10, %zmm10
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm12))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm9))
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = 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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,3,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = 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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6,7,8],ymm4[9],ymm2[10,11],ymm4[12],ymm2[13,14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm3))
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
+; AVX512-FCP-NEXT: vpermd %ymm7, %ymm3, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3,4],ymm0[5],ymm1[6,7,8,9],ymm0[10],ymm1[11,12],ymm0[13],ymm1[14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm0 = ymm3 ^ (mem & (ymm0 ^ ymm3))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm2))
+; AVX512-FCP-NEXT: vmovdqa %ymm0, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 64(%rax)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -3217,126 +3219,128 @@ define void @store_i16_stride7_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512DQ-FCP-LABEL: store_i16_stride7_vf16:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm2
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vporq %ymm0, %ymm3, %ymm16
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm12
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[u,u,u,u,u,u,14,15],zero,zero,ymm4[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm4[u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm14
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm2[16,17],zero,zero,zero,zero
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm1[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm1[u,u,u,u]
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm18
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,u,u,u,u,26,27,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3],ymm0[4],ymm3[5,6,7,8],ymm0[9],ymm3[10,11],ymm0[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm4[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [2,0,3,2,0,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm9
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-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-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm19
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm0[2],ymm3[3,4],ymm0[5],ymm3[6,7,8,9],ymm0[10],ymm3[11,12],ymm0[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm1[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-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-FCP-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-FCP-NEXT: vmovdqa (%rdx), %xmm8
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,2,3,3,10,0,11,10]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm13
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm12, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm3 = xmm11[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm3[0,1],xmm0[2],xmm3[3,4],xmm0[5],xmm3[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,0,1,0,0,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm15
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm14[1],xmm8[2,3],xmm14[4],xmm8[5,6],xmm14[7]
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm14
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm11, %zmm11
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm8
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm10, %xmm12
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm10
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10, %ymm10
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm8[0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [0,0,0,1,8,9,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm14, %zmm8, %zmm20
-; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-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-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm10, %zmm10
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0,1,2],ymm7[3],ymm5[4,5],ymm7[6],ymm5[7,8,9,10],ymm7[11],ymm5[12,13],ymm7[14],ymm5[15]
-; AVX512DQ-FCP-NEXT: 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-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
-; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpermd %zmm8, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm7 = [6,0,0,0,7,0,0,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm7, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm12[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,1,3,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,2,3]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4],ymm1[5],ymm2[6,7,8,9],ymm1[10],ymm2[11,12],ymm1[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
-; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rcx
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm17, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm18, %zmm8, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm2 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm19[2,2,2,3,6,6,6,7]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm9 ^ (mem & (zmm0 ^ zmm9))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm13))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm11[0,0,1,1,4,4,5,5]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm15 ^ (mem & (zmm0 ^ zmm15))
-; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm20))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm5))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm7 ^ (mem & (ymm1 ^ ymm7))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 ^ (mem & (ymm1 ^ ymm4))
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, 192(%rcx)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rcx)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rcx)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 64(%rcx)
+; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm1
+; AVX512DQ-FCP-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
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm9
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm6
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm10, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm12 = xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm12 = xmm12[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm12, %zmm14
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm1[16,17],zero,zero,zero,zero
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm0[u,u,u,u,u,u,u,u,12,13,14,15],zero,zero,ymm0[u,u,u,u,u,u,u,u,u,u,16,17],zero,zero,ymm0[u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm12
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm13
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm7 = xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm15, %xmm17
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm7, %zmm16
+; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%r10), %ymm6
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm15 = 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]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm16 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm14))
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm9, %xmm14
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm15 = xmm8[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0,1],xmm14[2],xmm15[3,4],xmm14[5],xmm15[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm15 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,0,1,0,0,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm14, %zmm15, %zmm16
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm10 = xmm10[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3],xmm10[4],xmm11[5,6],xmm10[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm8[0,0,1,1,4,4,5,5]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm16 ^ (mem & (zmm9 ^ zmm16))
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,16,17,18,19,24,25,26,27,30,31,28,29,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [0,0,0,1,2,3,3,0]
+; AVX512DQ-FCP-NEXT: vpermq %zmm8, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT: vpbroadcastd (%r10), %ymm8
+; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%r10), %ymm11
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm10))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = 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]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3],ymm9[4],ymm10[5,6,7,8],ymm9[9],ymm10[10,11],ymm9[12],ymm10[13,14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm4[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm2[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,0,3,2,0,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm9, %zmm10, %zmm11
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm2[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm3[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm10 = 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-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm11 ^ (mem & (zmm9 ^ zmm11))
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm1, %ymm10
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm0[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm0[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1,2],ymm11[3],ymm12[4,5],ymm11[6],ymm12[7,8,9,10],ymm11[11],ymm12[12,13],ymm11[14],ymm12[15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm12 = [2,2,3,3,10,0,11,10]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm10, %zmm11, %zmm12
+; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [0,13,4,0,0,14,5,0,0,13,4,0,0,14,5,0]
+; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpermd %zmm7, %zmm10, %zmm10
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm12))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (mem & (zmm10 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = 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]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,1,3,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = 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]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3],ymm4[4],ymm2[5,6,7,8],ymm4[9],ymm2[10,11],ymm4[12],ymm2[13,14,15]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm2 = ymm2 ^ (mem & (ymm2 ^ ymm3))
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,0,0,0,7,0,0,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,u,u,28,29,26,27,u,u,30,31,30,31]
+; AVX512DQ-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm0 = ymm3 ^ (mem & (ymm0 ^ ymm3))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm2))
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 64(%rax)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
@@ -6200,282 +6204,286 @@ define void @store_i16_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512-FCP-LABEL: store_i16_stride7_vf32:
; AVX512-FCP: # %bb.0:
; AVX512-FCP-NEXT: subq $296, %rsp # imm = 0x128
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm9
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm2
; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm4
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm22
; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm12
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm8, %ymm3
-; AVX512-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm5
-; AVX512-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpor %ymm3, %ymm5, %ymm3
-; AVX512-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm12, %ymm4
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm5
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm17
+; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm6
; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm15
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm3
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm11
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm11, %ymm7
-; AVX512-FCP-NEXT: vpor %ymm3, %ymm7, %ymm3
-; AVX512-FCP-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm2
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm12
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm3
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm5
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm7
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm9
+; AVX512-FCP-NEXT: vpor %ymm5, %ymm9, %ymm5
+; AVX512-FCP-NEXT: vmovdqu %ymm5, (%rsp) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm3
; AVX512-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm30
; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm21
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm24
-; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512-FCP-NEXT: vprold $16, %xmm14, %xmm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm27
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,0,0,1,8,8,9,0]
+; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm13[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1],xmm1[2,3],xmm0[4],xmm1[5,6],xmm0[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [8,0,9,0,0,0,1,1]
+; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm26
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm10, %ymm2
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = 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]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = 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]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm10, %ymm16
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0],ymm4[1],ymm8[2,3],ymm4[4],ymm8[5,6,7,8],ymm4[9],ymm8[10,11],ymm4[12],ymm8[13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,2,2,3,10,0,11,0]
+; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm4, %zmm23
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,8,0,9,0]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm24
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm14[0],xmm6[1],xmm14[1],xmm6[2],xmm14[2],xmm6[3],xmm14[3]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm0
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm21
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,0,8,8,9]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm3
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm5
-; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm8, %zmm23
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm4[0,1],xmm0[2],xmm4[3,4],xmm0[5],xmm4[6,7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm4
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm26
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [8,9,9,0,0,0,1,1]
-; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm28
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [8,0,9,0,0,0,1,1]
-; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm3, %zmm27
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm5
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm9
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm4
-; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,0]
-; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm29
-; AVX512-FCP-NEXT: vprold $16, %ymm15, %ymm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm6 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm8[3],ymm4[4,5],ymm8[6],ymm4[7,8,9,10],ymm8[11],ymm4[12,13],ymm8[14],ymm4[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,0,11,10]
-; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm4, %zmm31
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm20
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm3
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [2,1,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm28
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [2,2,2,3,0,8,8,9]
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm8
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm6
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm31
+; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm25
+; AVX512-FCP-NEXT: vprold $16, %ymm15, %ymm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
+; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} 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]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm4
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,0,2,1,4,4,6,5]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,2,3,3,10,0,11,10]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm29
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm13[4],xmm3[5],xmm13[5],xmm3[6],xmm13[6],xmm3[7],xmm13[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm7
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm0[4],xmm7[5],xmm0[5],xmm7[6],xmm0[6],xmm7[7],xmm0[7]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm17
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm1
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm10
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm30
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm12[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7,8,9],ymm8[10],ymm1[11,12],ymm8[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm1[2],ymm9[3,4],ymm1[5],ymm9[6,7,8,9],ymm1[10],ymm9[11,12],ymm1[13],ymm9[14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm13
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm6[3],ymm1[4,5],ymm6[6],ymm1[7,8,9,10],ymm6[11],ymm1[12,13],ymm6[14],ymm1[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm9
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm2
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512-FCP-NEXT: vpermd %ymm1, %ymm22, %ymm1
-; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = 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]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7,8,9],ymm12[10],ymm0[11,12],ymm12[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm3
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm4
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm27
+; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm3[0,1],ymm11[2],ymm3[3,4],ymm11[5],ymm3[6,7,8,9],ymm11[10],ymm3[11,12],ymm11[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm16
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm20[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm3[2],ymm11[3,4],ymm3[5],ymm11[6,7,8,9],ymm3[10],ymm11[11,12],ymm3[13],ymm11[14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm30, %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm3
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm20, %zmm5
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm1
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-FCP-NEXT: vpermd %ymm3, %ymm20, %ymm3
+; AVX512-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm3
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm9[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0],xmm1[1],xmm10[2,3],xmm1[4],xmm10[5,6],xmm1[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,1,1,8,8,0,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm9, %zmm10
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,3,3,0,0,0,1,1]
+; AVX512-FCP-NEXT: vpermq %zmm1, %zmm2, %zmm8
+; AVX512-FCP-NEXT: vprold $16, %xmm7, %xmm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2],xmm2[3,4],xmm1[5],xmm2[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3]
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm7
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm9, %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm19 ^ (zmm30 & (zmm10 ^ zmm19))
+; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm12
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm8))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm10))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm26 = zmm26 ^ (mem & (zmm26 ^ zmm0))
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm2
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,0,0,1,2,3,3,0]
+; AVX512-FCP-NEXT: vpermq %zmm0, %zmm2, %zmm31
+; AVX512-FCP-NEXT: vmovdqa %ymm14, %ymm9
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1],ymm7[2],ymm0[3,4],ymm7[5],ymm0[6,7,8,9],ymm7[10],ymm0[11,12],ymm7[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vprold $16, %ymm13, %ymm7
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm7[2],ymm14[3,4],ymm7[5],ymm14[6,7,8,9],ymm7[10],ymm14[11,12],ymm7[13],ymm14[14,15]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm15 = xmm13[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm15[0],xmm7[1],xmm15[2,3],xmm7[4],xmm15[5,6],xmm7[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,0,1,1,8,8,0,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm15, %zmm7
-; AVX512-FCP-NEXT: vprold $16, %xmm3, %xmm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm4[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm11, %xmm4
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm3, %xmm11
-; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm15, %zmm11
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm0, %zmm14
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm17[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1],ymm3[2],ymm6[3,4],ymm3[5],ymm6[6,7,8,9],ymm3[10],ymm6[11,12],ymm3[13],ymm6[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm0, %zmm3
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm12
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm24 ^ (zmm17 & (zmm12 ^ zmm24))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12))
-; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm5
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512-FCP-NEXT: vpermd %zmm5, %zmm6, %zmm6
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm14))
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [6,0,0,0,7,0,0,7]
-; AVX512-FCP-NEXT: vpermd %ymm10, %ymm12, %ymm10
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm26 ^ (zmm12 & (zmm7 ^ zmm26))
-; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm13
-; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm14
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm28))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm7))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm11))
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6,7,8],ymm2[9],ymm7[10,11],ymm2[12],ymm7[13,14,15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm21[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4],ymm8[5],ymm7[6,7,8,9],ymm8[10],ymm7[11,12],ymm8[13],ymm7[14,15]
+; AVX512-FCP-NEXT: vprold $16, %ymm15, %ymm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm21[1,2,2,3,5,6,6,7]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm10, %zmm8
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-FCP-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = 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]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm24[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm11[0],ymm7[1],ymm11[2,3],ymm7[4],ymm11[5,6,7,8],ymm7[9],ymm11[10,11],ymm7[12],ymm11[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm19[0,0,1,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm18[2,2,2,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm24[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm16, %zmm0
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512-FCP-NEXT: vpermd %zmm5, %zmm22, %zmm2
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm31))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm3 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm12 & (zmm3 ^ zmm0))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm25 ^ (zmm17 & (zmm21 ^ zmm25))
-; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm4, %zmm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm30
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm9 & mem)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm4))
-; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm3
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm3
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm23))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm21))
-; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm12
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm22[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm22[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7,8,9],ymm15[10],ymm14[11,12],ymm15[13],ymm14[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm10, %zmm14
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm7
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm10[2],ymm7[3,4],ymm10[5],ymm7[6,7,8,9],ymm10[10],ymm7[11,12],ymm10[13],ymm7[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm23 ^ (zmm2 & (zmm7 ^ zmm23))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
+; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm7
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm12 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512-FCP-NEXT: vpermd %zmm7, %zmm12, %zmm12
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm8))
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [6,0,0,0,7,0,0,7]
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm8, %ymm4
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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]
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm8[1],ymm15[2,3],ymm8[4],ymm15[5,6,7,8],ymm8[9],ymm15[10,11],ymm8[12],ymm15[13,14,15]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm18, %xmm9
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm17[0,0,1,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm16[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0,1],ymm0[2],ymm9[3,4],ymm0[5],ymm9[6,7,8,9],ymm0[10],ymm9[11,12],ymm0[13],ymm9[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm10, %zmm0
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm14))
+; AVX512-FCP-NEXT: vpermd %zmm7, %zmm20, %zmm7
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm29))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm0))
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm0 ^ (zmm30 & (zmm8 ^ zmm0))
+; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm6, %zmm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm27
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm8))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm6, %zmm6
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm6))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm5 & mem)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm8))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm28 = zmm24 ^ (zmm2 & (zmm28 ^ zmm24))
+; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm2
; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm31))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm26))
+; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm5
; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm29))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm27))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm25))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm28))
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 384(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 384(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 256(%rax)
; AVX512-FCP-NEXT: addq $296, %rsp # imm = 0x128
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
@@ -6823,282 +6831,286 @@ define void @store_i16_stride7_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
; AVX512DQ-FCP-LABEL: store_i16_stride7_vf32:
; AVX512DQ-FCP: # %bb.0:
; AVX512DQ-FCP-NEXT: subq $296, %rsp # imm = 0x128
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm9
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm9, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm17
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm22
; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm8
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm10
; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm12, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm6, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm17
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm6
; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm15
-; AVX512DQ-FCP-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]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm11, %ymm7
-; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm7, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm3, (%rsp) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm12
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm3
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm15, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,u,u,u,u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm9
+; AVX512DQ-FCP-NEXT: vpor %ymm5, %ymm9, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm5, (%rsp) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm11
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm3
; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm3, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm14, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm20
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm22
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm30
; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm21
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm12[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm13[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm12[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,10,0,11,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm24
-; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa 32(%rsi), %xmm14
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm14, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm6[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm27
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [0,0,0,1,8,8,9,0]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm13
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm13[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1],xmm1[2,3],xmm0[4],xmm1[5,6],xmm0[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [8,0,9,0,0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm26
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm10, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm11[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = 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-FCP-NEXT: vpshufb {{.*#+}} ymm4 = 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-FCP-NEXT: vmovdqa64 %ymm10, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm11[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm8[0],ymm4[1],ymm8[2,3],ymm4[4],ymm8[5,6,7,8],ymm4[9],ymm8[10,11],ymm4[12],ymm8[13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,2,2,3,10,0,11,0]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm4, %zmm23
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm10
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm22[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,2,2,3,8,0,9,0]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm24
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm14[0],xmm6[1],xmm14[1],xmm6[2],xmm14[2],xmm6[3],xmm14[3]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm12, %ymm1
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm27
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [0,2,2,3,8,0,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm25
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm10[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm0
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm21
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm11[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [2,2,2,3,0,8,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm2
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm5
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm8, %zmm23
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm4[0,1],xmm0[2],xmm4[3,4],xmm0[5],xmm4[6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm5, %xmm4
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,0,1,8,8,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm26
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [8,9,9,0,0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm28
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm0
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm0[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm5[0],xmm3[1],xmm5[2,3],xmm3[4],xmm5[5,6],xmm3[7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [8,0,9,0,0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm3, %zmm27
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm5
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm9
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm4, %xmm3, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm3, %xmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,0]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm29
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm11[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vpbroadcastd {{.*#+}} ymm6 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm11[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm8[3],ymm4[4,5],ymm8[6],ymm4[7,8,9,10],ymm8[11],ymm4[12,13],ymm8[14],ymm4[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm31 = [2,2,3,3,10,0,11,10]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm4, %zmm31
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm20
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm3
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm19
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm28 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm28
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm7[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm25 = [2,2,2,3,0,8,8,9]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm6
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm2, %xmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm31
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm0, %zmm25
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm15, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm7[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1],ymm0[2],ymm4[3,4],ymm0[5],ymm4[6,7,8,9],ymm0[10],ymm4[11,12],ymm0[13],ymm4[14,15]
+; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} 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-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm4
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,2,3,3,10,0,11,10]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm29
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm13[4],xmm3[5],xmm13[5],xmm3[6],xmm13[6],xmm3[7],xmm13[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm0, %xmm18
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm7
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm7[4],xmm0[4],xmm7[5],xmm0[5],xmm7[6],xmm0[6],xmm7[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm17
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm10
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm30
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm11 = xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm12[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1],ymm8[2],ymm1[3,4],ymm8[5],ymm1[6,7,8,9],ymm8[10],ymm1[11,12],ymm8[13],ymm1[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm18
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm1 = ymm7[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm14[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm1[2],ymm9[3,4],ymm1[5],ymm9[6,7,8,9],ymm1[10],ymm9[11,12],ymm1[13],ymm9[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm16[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm6[3],ymm1[4,5],ymm6[6],ymm1[7,8,9,10],ymm6[11],ymm1[12,13],ymm6[14],ymm1[15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm9
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm22, %ymm1
-; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = 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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm0[0,1],ymm12[2],ymm0[3,4],ymm12[5],ymm0[6,7,8,9],ymm12[10],ymm0[11,12],ymm12[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm3
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm27
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm3[0,1],ymm11[2],ymm3[3,4],ymm11[5],ymm3[6,7,8,9],ymm11[10],ymm3[11,12],ymm11[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm14[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm20, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm20[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm3[2],ymm11[3,4],ymm3[5],ymm11[6,7,8,9],ymm3[10],ymm11[11,12],ymm3[13],ymm11[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm30, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm20, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm3, %ymm1
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm3, %ymm20, %ymm3
+; AVX512DQ-FCP-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm3
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm9[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm9 = xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0],xmm1[1],xmm10[2,3],xmm1[4],xmm10[5,6],xmm1[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,1,1,8,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm9, %zmm10
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,3,3,0,0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermq %zmm1, %zmm2, %zmm8
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm7, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2],xmm2[3,4],xmm1[5],xmm2[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3]
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm0, %xmm0
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm9, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm30 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm19 ^ (zmm30 & (zmm10 ^ zmm19))
+; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm12
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm8))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm10))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm26 = zmm26 ^ (mem & (zmm26 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,0,0,1,2,3,3,0]
+; AVX512DQ-FCP-NEXT: vpermq %zmm0, %zmm2, %zmm31
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm14, %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm14[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm16[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm0[0,1],ymm7[2],ymm0[3,4],ymm7[5],ymm0[6,7,8,9],ymm7[10],ymm0[11,12],ymm7[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm13, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm16[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm7[2],ymm14[3,4],ymm7[5],ymm14[6,7,8,9],ymm7[10],ymm14[11,12],ymm7[13],ymm14[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm0[u,u,4,5,u,u,u,u,6,7,u,u,u,u,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm13[4],xmm0[4],xmm13[5],xmm0[5],xmm13[6],xmm0[6],xmm13[7],xmm0[7]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm15 = xmm13[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm7 = xmm15[0],xmm7[1],xmm15[2,3],xmm7[4],xmm15[5,6],xmm7[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,0,1,1,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm15, %zmm7
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm3, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm4[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm13[0,1],xmm0[2],xmm13[3,4],xmm0[5],xmm13[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm11, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm3, %xmm11
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm15, %zmm11
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm0 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm0, %zmm14
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm17[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-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-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm13[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm17[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm6[0,1],ymm3[2],ymm6[3,4],ymm3[5],ymm6[6,7,8,9],ymm3[10],ymm6[11,12],ymm3[13],ymm6[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm0, %zmm3
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm12
-; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm24 ^ (zmm17 & (zmm12 ^ zmm24))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12))
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm5
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm6, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm14))
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [6,0,0,0,7,0,0,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm10, %ymm12, %ymm10
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm26 ^ (zmm12 & (zmm7 ^ zmm26))
-; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm13
-; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm14
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm28))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm7))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm27 = zmm27 ^ (mem & (zmm27 ^ zmm11))
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6,7,8],ymm2[9],ymm7[10,11],ymm2[12],ymm7[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm21[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm8[2],ymm7[3,4],ymm8[5],ymm7[6,7,8,9],ymm8[10],ymm7[11,12],ymm8[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm15, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm21[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm10, %zmm8
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-FCP-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = 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-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm24[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm11[0],ymm7[1],ymm11[2,3],ymm7[4],ymm11[5,6,7,8],ymm7[9],ymm11[10,11],ymm7[12],ymm11[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm2
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm19[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm18[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm24[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm2[2],ymm0[3,4],ymm2[5],ymm0[6,7,8,9],ymm2[10],ymm0[11,12],ymm2[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm16, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vpermd %zmm5, %zmm22, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm31))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm3 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm12 & (zmm3 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm25 ^ (zmm17 & (zmm21 ^ zmm25))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm4, %zmm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm30
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm9 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm3
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm10, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm23))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm21))
-; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm4
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm12
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm22[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4],ymm15[5],ymm12[6,7,8,9],ymm15[10],ymm12[11,12],ymm15[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm15 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm22[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4],ymm15[5],ymm14[6,7,8,9],ymm15[10],ymm14[11,12],ymm15[13],ymm14[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm10, %zmm14
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm7
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm10 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm10[2],ymm7[3,4],ymm10[5],ymm7[6,7,8,9],ymm10[10],ymm7[11,12],ymm10[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm10 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm23 ^ (zmm2 & (zmm7 ^ zmm23))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm7))
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm7
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm12 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512DQ-FCP-NEXT: vpermd %zmm7, %zmm12, %zmm12
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm8))
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm8 = [6,0,0,0,7,0,0,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm8, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm16[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm15[0],ymm8[1],ymm15[2,3],ymm8[4],ymm15[5,6,7,8],ymm8[9],ymm15[10,11],ymm8[12],ymm15[13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm18, %xmm9
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm15 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm19 = ymm17[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm16[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0,1],ymm0[2],ymm9[3,4],ymm0[5],ymm9[6,7,8,9],ymm0[10],ymm9[11,12],ymm0[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm10, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm14))
+; AVX512DQ-FCP-NEXT: vpermd %zmm7, %zmm20, %zmm7
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm29))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (mem & (zmm7 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm15, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm19, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm0 ^ (zmm30 & (zmm8 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsp), %zmm6, %zmm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = (zmm0 & mem) | zmm27
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm8))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm6, %zmm6
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (mem & (zmm8 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 | (zmm5 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm8))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm28 = zmm24 ^ (zmm2 & (zmm28 ^ zmm24))
+; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm2
; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm5
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm31))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm26))
+; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm5
; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm29))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm27))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm25))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm28))
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 320(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 384(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 128(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 384(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 256(%rax)
; AVX512DQ-FCP-NEXT: addq $296, %rsp # imm = 0x128
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
@@ -13225,627 +13237,642 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512-FCP-LABEL: store_i16_stride7_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: subq $1480, %rsp # imm = 0x5C8
-; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm3
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm0
+; AVX512-FCP-NEXT: subq $1768, %rsp # imm = 0x6E8
+; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm2
; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %ymm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm2
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm4
+; AVX512-FCP-NEXT: vpor %ymm2, %ymm4, %ymm4
; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %ymm2
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm4
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm4, %ymm5
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm9, %ymm25
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm7
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %ymm5
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm10
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm10, %ymm11
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa 64(%r9), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm7
; AVX512-FCP-NEXT: vmovdqa 64(%r8), %ymm13
+; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm9
-; AVX512-FCP-NEXT: vmovdqa64 %ymm13, %ymm26
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm9, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm27
-; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
-; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
+; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm13
; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm14
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm13
; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm14
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm7
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm13
; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm13
; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm15
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm13
-; AVX512-FCP-NEXT: vpor %ymm15, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm13
; AVX512-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm6
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm14, %ymm8
-; AVX512-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
-; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm6
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm9, %ymm8
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm25
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm23
+; AVX512-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm3
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm22
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm6
+; AVX512-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
+; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm6
+; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm16
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm7
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm8
; AVX512-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
-; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm8
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm8, %ymm6
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm10
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm11
-; AVX512-FCP-NEXT: vporq %ymm11, %ymm6, %ymm21
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm12
-; AVX512-FCP-NEXT: vmovdqa64 %ymm13, %ymm17
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm7[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm11
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm12
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm11 ^ (mem & (zmm12 ^ zmm11))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm5
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u],zero,zero,zero,zero,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
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm11 = ymm11 | (ymm12 & ymm13)
-; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm6
-; AVX512-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 | (ymm11 & ymm16)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm20
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm12, %ymm11
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
-; AVX512-FCP-NEXT: vpermd %ymm5, %ymm12, %ymm12
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm16 & (ymm12 ^ ymm11))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
-; AVX512-FCP-NEXT: vprold $16, %ymm6, %ymm11
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm11 = ymm11 ^ (ymm16 & (ymm11 ^ ymm12))
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm11[0,1,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0,1],ymm12[2],ymm9[3,4],ymm12[5],ymm9[6,7,8,9],ymm12[10],ymm9[11,12],ymm12[13],ymm9[14,15]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm11, %zmm8
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm4 ^ (mem & (zmm8 ^ zmm4))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %ymm4
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[16,17,u,u,u,u],zero,zero
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm9 = ymm9 | (ymm8 & ymm11)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
+; AVX512-FCP-NEXT: vmovdqa 96(%r9), %ymm12
+; AVX512-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm12, %ymm10
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm10 = ymm10 | (ymm9 & ymm11)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm21
+; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm8
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,4,0,0,0,5,0,0]
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm9, %ymm9
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm9 = ymm9 ^ (ymm3 & (ymm9 ^ ymm8))
+; AVX512-FCP-NEXT: vprold $16, %ymm12, %ymm8
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm8 ^ (ymm19 & (ymm8 ^ ymm9))
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm10[0,1,2,3],zmm8[0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
-; AVX512-FCP-NEXT: vmovdqa 96(%rax), %ymm4
-; AVX512-FCP-NEXT: vpermd %ymm4, %ymm11, %ymm11
-; AVX512-FCP-NEXT: vpandn %ymm11, %ymm13, %ymm11
-; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm12
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm4
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm11[0,1,2],ymm4[3],ymm11[4,5],ymm4[6],ymm11[7,8,9,10],ymm4[11],ymm11[12,13],ymm4[14],ymm11[15]
-; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm11 = [151522058,0,421010202,421010202]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm11, %ymm24
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6,7,8],ymm2[9],ymm7[10,11],ymm2[12],ymm7[13,14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,2,2,3,10,9,11,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm2
-; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm11 = [218894094,0,488382238,488382238]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm4
-; AVX512-FCP-NEXT: vmovdqa64 %ymm11, %ymm16
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm11[0],ymm4[1],ymm11[2,3],ymm4[4],ymm11[5,6,7,8],ymm4[9],ymm11[10,11],ymm4[12],ymm11[13,14,15]
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,1,4,5,4,5,5,7]
+; AVX512-FCP-NEXT: vmovdqa 96(%rax), %ymm8
+; AVX512-FCP-NEXT: vpermd %ymm8, %ymm9, %ymm9
+; AVX512-FCP-NEXT: vpandn %ymm9, %ymm13, %ymm9
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm8
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5],ymm8[6],ymm9[7,8,9,10],ymm8[11],ymm9[12,13],ymm8[14],ymm9[15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,2,2,3,10,9,11,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm26, %zmm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3],ymm5[4],ymm8[5,6,7,8],ymm5[9],ymm8[10,11],ymm5[12],ymm8[13,14,15]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,2,2,3,8,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm1
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm2 ^ (zmm23 & (zmm1 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [5,0,0,0,6,0,0,6]
-; AVX512-FCP-NEXT: vpermd %ymm5, %ymm1, %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT: vpbroadcastd 72(%rax), %ymm1
-; AVX512-FCP-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512-FCP-NEXT: vmovdqa 64(%rax), %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm1
-; AVX512-FCP-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm2
-; AVX512-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm20, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm2 ^ (zmm28 & (zmm1 ^ zmm2))
; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3],ymm1[4,5],ymm2[6],ymm1[7,8,9,10],ymm2[11],ymm1[12,13],ymm2[14],ymm1[15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,0,3,10,10,11,11]
-; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm2
-; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [5,0,0,0,6,0,0,6]
+; AVX512-FCP-NEXT: vpermd %ymm4, %ymm0, %ymm0
; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm13
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm2
-; AVX512-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm1
-; AVX512-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vmovdqa %ymm11, %ymm7
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm15, %ymm3
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm23 & (zmm0 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4],ymm2[5],ymm1[6,7,8,9],ymm2[10],ymm1[11,12],ymm2[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vprold $16, %ymm10, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm8[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm16
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512-FCP-NEXT: vpermd %zmm16, %zmm0, %zmm0
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm15
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm15, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm24, %ymm21
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm1
-; AVX512-FCP-NEXT: vmovdqa %ymm7, %ymm10
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm29[2,2,2,2,6,6,6,6]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm7
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm29[0,1,1,3,4,5,5,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = 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]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm1 & (zmm4 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm13
-; AVX512-FCP-NEXT: vprold $16, %ymm25, %ymm2
-; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm25
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm26[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm13, %ymm3
-; AVX512-FCP-NEXT: vmovdqa64 %ymm13, %ymm18
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm26[0,0,2,1,4,4,6,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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]
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,2,3,3,10,9,11,10]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
-; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm0
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512-FCP-NEXT: vpbroadcastd 72(%rax), %ymm0
+; AVX512-FCP-NEXT: vpandn %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa 64(%rax), %ymm1
+; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm12 & (zmm0 ^ zmm3))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm4))
+; AVX512-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
+; AVX512-FCP-NEXT: vpandn %ymm0, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa (%rax), %ymm1
+; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm29[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm0[0],ymm2[1],ymm0[2,3],ymm2[4],ymm0[5,6,7,8],ymm2[9],ymm0[10,11],ymm2[12],ymm0[13,14,15]
-; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm4
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm12 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm2
-; AVX512-FCP-NEXT: vmovdqa %xmm4, %xmm5
+; AVX512-FCP-NEXT: vpbroadcastd {{.*#+}} ymm1 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm18
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm29
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,0,3,10,10,11,11]
+; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm1
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm4
; AVX512-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2,3],xmm2[4],xmm4[5,6],xmm2[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [8,9,9,0,0,0,1,1]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm24, %zmm2
-; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm3, (%rsp) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm6
+; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm5
+; AVX512-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vprold $16, %xmm5, %xmm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,0,0,1,8,8,9,0]
+; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm0
+; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %xmm6
; AVX512-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vprold $16, %xmm6, %xmm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,1,1,8,8,0,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm26, %zmm5
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm4 & (zmm5 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm2
-; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm6
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm29, %zmm2
-; AVX512-FCP-NEXT: vpbroadcastd 64(%rax), %ymm6
-; AVX512-FCP-NEXT: vpbroadcastd 68(%rax), %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm30
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm20 & (zmm30 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm2 & (zmm30 ^ zmm5))
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa %xmm7, (%rsp) # 16-byte Spill
+; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm7[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm1[1],xmm5[2,3],xmm1[4],xmm5[5,6],xmm1[7]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,0,1,1,8,8,0,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm5
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm0 ^ (zmm24 & (zmm5 ^ zmm0))
+; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm6
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm6
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,3,3,0,0,0,1,1]
+; AVX512-FCP-NEXT: vpermq %zmm6, %zmm14, %zmm6
+; AVX512-FCP-NEXT: vpbroadcastd 100(%rax), %ymm8
+; AVX512-FCP-NEXT: vpbroadcastd 104(%rax), %ymm9
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm0
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm19 & (zmm0 ^ zmm6))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm17 & (zmm0 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %xmm8
; AVX512-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm5
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
+; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %xmm9
+; AVX512-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm5
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[1,1,2,2]
; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm24, %zmm5
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [8,9,9,0,0,0,1,1]
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm8, %zmm5
+; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm11
+; AVX512-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm12
; AVX512-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-FCP-NEXT: vprold $16, %xmm12, %xmm6
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm3[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm8, %xmm8
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm26, %zmm8
-; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm24
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm4 & (zmm8 ^ zmm5))
-; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm4
-; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm5
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm9
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm5 ^ (zmm6 & (zmm9 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqa 64(%r9), %xmm5
+; AVX512-FCP-NEXT: vmovdqa 64(%r8), %xmm11
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm5[4],xmm11[5],xmm5[5],xmm11[6],xmm5[6],xmm11[7],xmm5[7]
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm5[0],xmm11[1],xmm5[1],xmm11[2],xmm5[2],xmm11[3],xmm5[3]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm5
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm12 = [0,0,0,1,2,3,3,0]
+; AVX512-FCP-NEXT: vpermq %zmm5, %zmm12, %zmm5
+; AVX512-FCP-NEXT: vpbroadcastd 64(%rax), %ymm11
+; AVX512-FCP-NEXT: vpbroadcastd 68(%rax), %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm0
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm21 & (zmm0 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm5 & (zmm0 ^ zmm9))
+; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm11, %xmm9
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm13[0],xmm9[1],xmm13[2,3],xmm9[4],xmm13[5,6],xmm9[7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm13
+; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm8, %zmm9
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vprold $16, %xmm1, %xmm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm13[0,1],xmm8[2],xmm13[3,4],xmm8[5],xmm13[6,7]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm13
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm7, %zmm13
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm9 ^ (zmm6 & (zmm13 ^ zmm9))
+; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm6
+; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm8
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm6, %ymm6
+; AVX512-FCP-NEXT: vpermq %zmm6, %zmm12, %zmm6
+; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm8
+; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm9
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm1
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm21 & (zmm1 ^ zmm6))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm5 & (zmm1 ^ zmm13))
+; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512-FCP-NEXT: vprold $16, %xmm2, %xmm5
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,3]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm4, %zmm5
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm2
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm8[1,1,2,2]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2,3],xmm2[4],xmm4[5,6],xmm2[7]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm2
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm5 ^ (zmm24 & (zmm2 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm5
-; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm4
-; AVX512-FCP-NEXT: vmovdqa64 %xmm7, %xmm26
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm29, %zmm4
-; AVX512-FCP-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512-FCP-NEXT: vpbroadcastd 4(%rax), %ymm6
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm29
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (zmm20 & (zmm29 ^ zmm4))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (zmm2 & (zmm29 ^ zmm8))
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
-; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm4
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4],ymm5[5],ymm4[6,7,8,9],ymm5[10],ymm4[11,12],ymm5[13],ymm4[14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm4
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm20[2,2,2,2,6,6,6,6]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm20[0,1,1,3,4,5,5,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm2
-; AVX512-FCP-NEXT: vmovdqa 96(%rcx), %xmm12
-; AVX512-FCP-NEXT: vmovdqa 96(%rdx), %xmm11
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm5, %xmm5
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,2,2,3,8,9,9,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm9, %zmm13
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm5
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm5
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm28[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1,2],ymm5[3],ymm6[4,5],ymm5[6],ymm6[7,8,9,10],ymm5[11],ymm6[12,13],ymm5[14],ymm6[15]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm4 ^ (zmm1 & (zmm2 ^ zmm4))
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512-FCP-NEXT: vprold $16, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm17[1,2,2,3,5,6,6,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm5, %ymm28
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[0,0,2,1,4,4,6,5]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm4
-; AVX512-FCP-NEXT: vmovdqa 96(%rsi), %xmm1
-; AVX512-FCP-NEXT: vmovdqa 96(%rdi), %xmm5
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,1,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm19, %zmm8
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm13 ^ (zmm23 & (zmm8 ^ zmm13))
-; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm22 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512-FCP-NEXT: # ymm22 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vpermq %zmm1, %zmm14, %zmm0
+; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm30
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm19 & (zmm30 ^ zmm0))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm17 & (zmm30 ^ zmm2))
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm2
+; AVX512-FCP-NEXT: vmovdqu64 %ymm22, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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]
+; AVX512-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm6 = [151522058,0,421010202,421010202]
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm15[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm3
+; AVX512-FCP-NEXT: vmovdqu64 %ymm25, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm22 = [218894094,0,488382238,488382238]
; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm0
-; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm6
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512-FCP-NEXT: vmovdqu64 %ymm23, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm23[3,3,3,3,7,7,7,7]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm2
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm23[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm2
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm1 ^ (zmm28 & (zmm2 ^ zmm1))
+; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm2
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vprold $16, %ymm18, %ymm1
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [2,1,3,2,10,10,10,11]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm1
+; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm17
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512-FCP-NEXT: vpermd %zmm17, %zmm0, %zmm21
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm21 ^ (mem & (zmm21 ^ zmm1))
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm19
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm24[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3],ymm0[4],ymm1[5,6,7,8],ymm0[9],ymm1[10,11],ymm0[12],ymm1[13,14,15]
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm1
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm25
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7,8,9],ymm4[10],ymm0[11,12],ymm4[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm11[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7,8,9],ymm4[10],ymm7[11,12],ymm4[13],ymm7[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm7
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm1 ^ (zmm18 & (zmm7 ^ zmm1))
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512-FCP-NEXT: vprold $16, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0,1,2],ymm1[3],ymm0[4,5],ymm1[6],ymm0[7,8,9,10],ymm1[11],ymm0[12,13],ymm1[14],ymm0[15]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm25[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
-; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
-; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
-; AVX512-FCP-NEXT: vpermd %zmm16, %zmm0, %zmm16
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm4))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm2))
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm0
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm2
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm21[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm3[0,1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7,8,9,10],ymm2[11],ymm3[12,13],ymm2[14],ymm3[15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vmovdqa64 %xmm15, %xmm27
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm14
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,2,2,3,8,8,8,9]
-; AVX512-FCP-NEXT: vmovdqa 96(%r9), %xmm4
-; AVX512-FCP-NEXT: vmovdqa 96(%r8), %xmm3
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm13
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm15
-; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm0 ^ (zmm23 & (zmm14 ^ zmm0))
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [6,7,3,3,7,7,6,7]
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm18, %ymm0 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpbroadcastd 96(%rax), %ymm19
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm0, %zmm19
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm23 & (zmm19 ^ zmm6))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm26 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm26 & (zmm19 ^ zmm8))
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm1, %xmm0
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm5[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
-; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
-; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm5
-; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm5[2],ymm1[3,4],ymm5[5],ymm1[6,7,8,9],ymm5[10],ymm1[11,12],ymm5[13],ymm1[14,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm8
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm5, %xmm15
-; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm1
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,0,1,8,8,9,0]
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm0
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm20 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm15
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm12, %xmm12
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm12 = xmm11[0],xmm12[1],xmm11[2,3],xmm12[4],xmm11[5,6],xmm12[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vmovdqa64 %zmm24, %zmm25
-; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm24, %zmm12
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm0 ^ (zmm28 & (zmm12 ^ zmm0))
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm22 = ymm22 ^ (ymm23 & (ymm22 ^ ymm2))
-; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm2, %ymm2
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm3
-; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm15
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm15[0],ymm0[1],ymm15[2,3],ymm0[4],ymm15[5,6,7,8],ymm0[9],ymm15[10,11],ymm0[12],ymm15[13,14,15]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm15
+; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm31
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[1,1,1,1,5,5,5,5]
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm13
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm20[2,2,2,2,6,6,6,6]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm0[0,1],ymm15[2],ymm0[3,4],ymm15[5],ymm0[6,7,8,9],ymm15[10],ymm0[11,12],ymm15[13],ymm0[14,15]
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm20[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0,1],ymm0[2],ymm14[3,4],ymm0[5],ymm14[6,7,8,9],ymm0[10],ymm14[11,12],ymm0[13],ymm14[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm9, %zmm0
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,2,3,3,10,9,11,10]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm14, %zmm1
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm13 ^ (zmm18 & (zmm0 ^ zmm13))
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512-FCP-NEXT: vprold $16, %ymm9, %ymm4
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm27[1,2,2,3,5,6,6,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm4[2],ymm13[3,4],ymm4[5],ymm13[6,7,8,9],ymm4[10],ymm13[11,12],ymm4[13],ymm13[14,15]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm9, %ymm4
+; AVX512-FCP-NEXT: vmovdqa64 %ymm9, %ymm25
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm27[0,0,2,1,4,4,6,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm15[3],ymm4[4,5],ymm15[6],ymm4[7,8,9,10],ymm15[11],ymm4[12,13],ymm15[14],ymm4[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm14, %zmm4
+; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm13
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512-FCP-NEXT: vpermd %zmm13, %zmm14, %zmm18
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (zmm28 & (zmm18 ^ zmm1))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (zmm1 & (zmm18 ^ zmm7))
+; AVX512-FCP-NEXT: vmovdqa (%rsp), %xmm3 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm7 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm7 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm9
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm11, %ymm14
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm15[0],ymm14[1],ymm15[2,3],ymm14[4],ymm15[5,6,7,8],ymm14[9],ymm15[10,11],ymm14[12],ymm15[13,14,15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,2,2,3,8,9,9,0]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm22, %zmm14
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm7 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm7 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm15
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm24[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1,2],ymm15[3],ymm12[4,5],ymm15[6],ymm12[7,8,9,10],ymm15[11],ymm12[12,13],ymm15[14],ymm12[15]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [2,1,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm12
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm14 ^ (zmm23 & (zmm12 ^ zmm14))
+; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm7
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm14[2],ymm7[3,4],ymm14[5],ymm7[6,7,8,9],ymm14[10],ymm7[11,12],ymm14[13],ymm7[14,15]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm14
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,2,2,3,8,8,8,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm7
+; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]
+; AVX512-FCP-NEXT: vpermd %zmm17, %zmm14, %zmm16
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (zmm28 & (zmm16 ^ zmm4))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (zmm1 & (zmm16 ^ zmm0))
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [6,7,3,3,7,7,6,7]
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm1 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpbroadcastd 96(%rax), %ymm4
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm17
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (zmm4 & (zmm17 ^ zmm7))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (zmm19 & (zmm17 ^ zmm12))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (ymm4 & (ymm3 ^ ymm0))
+; AVX512-FCP-NEXT: vextracti64x4 $1, %zmm0, %ymm7
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufhw {{.*#+}} ymm10 = ymm0[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
+; AVX512-FCP-NEXT: vmovdqa %ymm11, %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm14
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm11
+; AVX512-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm0
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm20[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3],ymm0[4],ymm9[5,6,7,8],ymm0[9],ymm9[10,11],ymm0[12],ymm9[13,14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm0
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm8 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3]
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512-FCP-NEXT: vpshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm6 = mem[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm9[3],ymm6[4,5],ymm9[6],ymm6[7,8,9,10],ymm9[11],ymm6[12,13],ymm9[14],ymm6[15]
+; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm6
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm0 ^ (zmm23 & (zmm6 ^ zmm0))
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [8,9,9,0,0,0,1,1]
-; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm0
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufhw {{.*#+}} ymm3 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm7[0,0,1,1]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
-; AVX512-FCP-NEXT: vmovdqa64 %xmm15, %xmm31
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm7[0,0,1,1]
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm18, %ymm7 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm15
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm23 & (zmm7 ^ zmm1))
-; AVX512-FCP-NEXT: vpbroadcastd 100(%rax), %ymm1
-; AVX512-FCP-NEXT: vpbroadcastd 104(%rax), %ymm15
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm1, %zmm15
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm18 & (zmm15 ^ zmm0))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm26 & (zmm7 ^ zmm14))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm0 & (zmm15 ^ zmm12))
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512-FCP-NEXT: vprold $16, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,3]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0,1],xmm9[2],xmm10[3,4],xmm9[5],xmm10[6,7]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm9
-; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm12[0],xmm1[1],xmm12[2,3],xmm1[4],xmm12[5,6],xmm1[7]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm1
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm9 ^ (zmm28 & (zmm1 ^ zmm9))
-; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm9
-; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm5, %xmm5
-; AVX512-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm8
-; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm6
-; AVX512-FCP-NEXT: vpbroadcastd 36(%rax), %ymm5
-; AVX512-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm18 & (zmm5 ^ zmm6))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm0 & (zmm5 ^ zmm1))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogd {{.*#+}} ymm2 = m32bcst ^ (ymm0 & (ymm2 ^ m32bcst))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (ymm0 & (ymm3 ^ ymm2))
+; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm25, %ymm5
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm8
+; AVX512-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm27[3,3,3,3,7,7,7,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7,8,9],ymm9[10],ymm8[11,12],ymm9[13],ymm8[14,15]
+; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm8
+; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpbroadcastd 32(%rax), %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm8))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm19 & (zmm0 ^ zmm6))
+; AVX512-FCP-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]
+; AVX512-FCP-NEXT: vpternlogd {{.*#+}} ymm7 = m32bcst ^ (ymm1 & (ymm7 ^ m32bcst))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm10 = ymm10 ^ (ymm1 & (ymm10 ^ ymm7))
; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (ymm6 & (ymm4 ^ ymm22))
-; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[0,1,2,3],zmm3[0,1,2,3]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm1 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm1 = (zmm1 & zmm6) | mem
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm3 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm3 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm3 = (zmm3 & zmm6) | mem
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd (%rsp), %xmm6, %xmm6 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm6 & (ymm12 ^ ymm3))
+; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm12[0,1,2,3],zmm10[0,1,2,3]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512-FCP-NEXT: vpermd %zmm13, %zmm3, %zmm4
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm2))
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm2 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm2 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm2 = (zmm2 & zmm6) | mem
+; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm7
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm6 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm6 = (zmm6 & zmm7) | mem
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512-FCP-NEXT: # xmm8 = xmm8[4],mem[4],xmm8[5],mem[5],xmm8[6],mem[6],xmm8[7],mem[7]
-; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
-; AVX512-FCP-NEXT: # xmm9 = xmm9[4],mem[4],xmm9[5],mem[5],xmm9[6],mem[6],xmm9[7],mem[7]
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm6
-; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm9, %xmm9
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm10[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm11 = mem[0,1,1,3,4,5,5,7]
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: # ymm12 = mem[1,1,1,1,5,5,5,5]
-; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm11[0,1],ymm12[2],ymm11[3,4],ymm12[5],ymm11[6,7,8,9],ymm12[10],ymm11[11,12],ymm12[13],ymm11[14,15]
-; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm12
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,1,4,5,4,5,5,7]
-; AVX512-FCP-NEXT: vpermd %ymm12, %ymm13, %ymm13
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm12[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[16,17],zero,zero
-; AVX512-FCP-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]
-; AVX512-FCP-NEXT: vpandn %ymm13, %ymm14, %ymm13
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
-; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm13))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm4 ^ (zmm28 & (zmm6 ^ zmm4))
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm8 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm4 ^ (zmm28 & (zmm8 ^ zmm4))
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm1 ^ (zmm4 & (zmm6 ^ zmm1))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm3 ^ (zmm4 & (zmm8 ^ zmm3))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm9 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
+; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm10 # 16-byte Folded Reload
+; AVX512-FCP-NEXT: # xmm10 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm8
+; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm10, %xmm10
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm12 = mem[0,1,1,3,4,5,5,7]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: # ymm13 = mem[1,1,1,1,5,5,5,5]
+; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
+; AVX512-FCP-NEXT: vmovdqa 32(%rax), %ymm13
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [0,1,4,5,4,5,5,7]
+; AVX512-FCP-NEXT: vpermd %ymm13, %ymm14, %ymm14
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm13[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[16,17],zero,zero
+; AVX512-FCP-NEXT: vpandnq %ymm14, %ymm28, %ymm14
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
+; AVX512-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,3]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm1 ^ (mem & (zmm3 ^ zmm1))
-; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm1 = zmm1 | (zmm0 & mem)
-; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
-; AVX512-FCP-NEXT: # zmm12 = zmm12 | (zmm0 & mem)
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm3))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm21 ^ (mem & (zmm21 ^ zmm3))
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm7 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm5 ^ (zmm14 & (zmm7 ^ zmm5))
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm5 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm8 # 32-byte Folded Reload
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm14 & (zmm8 ^ zmm5))
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm2 ^ (zmm5 & (zmm7 ^ zmm2))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm5 & (zmm8 ^ zmm6))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm5
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (mem & (zmm5 ^ zmm2))
+; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm2 = zmm2 | (zmm1 & mem)
+; AVX512-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm13 # 64-byte Folded Reload
+; AVX512-FCP-NEXT: # zmm13 = zmm13 | (zmm1 & mem)
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm5))
; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 320(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 320(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 128(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm29, (%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 448(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 704(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 640(%rax)
-; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512-FCP-NEXT: vmovaps %zmm0, 576(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 640(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 384(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 256(%rax)
; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 512(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 832(%rax)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 768(%rax)
-; AVX512-FCP-NEXT: addq $1480, %rsp # imm = 0x5C8
+; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovaps %zmm0, (%rax)
+; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovaps %zmm0, 448(%rax)
+; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512-FCP-NEXT: vmovaps %zmm0, 704(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 512(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 832(%rax)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
+; AVX512-FCP-NEXT: addq $1768, %rsp # imm = 0x6E8
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -14552,627 +14579,642 @@ define void @store_i16_stride7_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.ve
;
; AVX512DQ-FCP-LABEL: store_i16_stride7_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: subq $1480, %rsp # imm = 0x5C8
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: subq $1768, %rsp # imm = 0x6E8
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm2
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm2, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm4
+; AVX512DQ-FCP-NEXT: vpor %ymm2, %ymm4, %ymm4
; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,13,14,15,128,128,u,u,u,u,u,u,u,u,u,u,16,17,128,128,u,u,u,u,u,u,u,u,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm4
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm4, %ymm5
-; AVX512DQ-FCP-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]
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm9, %ymm25
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128,128,128,128,128,128,128,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %ymm5
+; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm10
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm10, %ymm11
+; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa 64(%r9), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm7
; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm13, %ymm26
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm9, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm27
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm29
-; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm13
; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm22
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm0, %ymm28
-; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm13
; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm13
; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm14
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vpor %ymm14, %ymm15, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm13
; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm15
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm13
-; AVX512DQ-FCP-NEXT: vpor %ymm15, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm13
; AVX512DQ-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm14
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm14, %ymm8
-; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm9, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpor %ymm7, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm25
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm23
+; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm22
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm6
+; AVX512DQ-FCP-NEXT: vpor %ymm3, %ymm6, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm6
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm16
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm7
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm8
; AVX512DQ-FCP-NEXT: vpor %ymm6, %ymm8, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm8
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm8, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm10
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm11
-; AVX512DQ-FCP-NEXT: vporq %ymm11, %ymm6, %ymm21
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm1[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm12
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm13, %ymm17
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm7[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm16, %zmm11
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm12
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm11 ^ (mem & (zmm12 ^ zmm11))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm5[u,u],zero,zero,zero,zero,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
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm11 = ymm11 | (ymm12 & ymm13)
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 | (ymm11 & ymm16)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm20
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm12, %ymm11
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm12 = [0,4,0,0,0,5,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %ymm5, %ymm12, %ymm12
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm16 & (ymm12 ^ ymm11))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm6, %ymm11
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm16 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm11 = ymm11 ^ (ymm16 & (ymm11 ^ ymm12))
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm11[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm8 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm1[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4],ymm8[5],ymm9[6,7,8,9],ymm8[10],ymm9[11,12],ymm8[13],ymm9[14,15]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm5[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm11, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm4 ^ (mem & (zmm8 ^ zmm4))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %ymm4
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm4[u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[14,15,u,u,u,u],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[16,17,u,u,u,u],zero,zero
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm9 = ymm9 | (ymm8 & ymm11)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %ymm12
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm12, %ymm10
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm11 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm10 = ymm10 | (ymm9 & ymm11)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm21
+; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm8, %ymm8
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,4,0,0,0,5,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm9 = ymm9 ^ (ymm3 & (ymm9 ^ ymm8))
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm12, %ymm8
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm8 = ymm8 ^ (ymm19 & (ymm8 ^ ymm9))
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm10[0,1,2,3],zmm8[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rax), %ymm4
-; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm11, %ymm11
-; AVX512DQ-FCP-NEXT: vpandn %ymm11, %ymm13, %ymm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm12
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm4
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm7[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm11[0,1,2],ymm4[3],ymm11[4,5],ymm4[6],ymm11[7,8,9,10],ymm4[11],ymm11[12,13],ymm4[14],ymm11[15]
-; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm11 = [151522058,0,421010202,421010202]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm11, %ymm24
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm7[0],ymm2[1],ymm7[2,3],ymm2[4],ymm7[5,6,7,8],ymm2[9],ymm7[10,11],ymm2[12],ymm7[13,14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,2,2,3,10,9,11,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm2
-; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm11 = [218894094,0,488382238,488382238]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm11, %ymm16
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm11 = ymm1[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm11[0],ymm4[1],ymm11[2,3],ymm4[4],ymm11[5,6,7,8],ymm4[9],ymm11[10,11],ymm4[12],ymm11[13,14,15]
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
-; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rax), %ymm8
+; AVX512DQ-FCP-NEXT: vpermd %ymm8, %ymm9, %ymm9
+; AVX512DQ-FCP-NEXT: vpandn %ymm9, %ymm13, %ymm9
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [12,13,128,128,128,128,128,128,128,128,128,128,128,128,14,15,128,128,128,128,128,128,128,128,128,128,128,128,16,17,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm8
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm8 = 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-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm5[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5],ymm8[6],ymm9[7,8,9,10],ymm8[11],ymm9[12,13],ymm8[14],ymm9[15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[10,11,8,9,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,u,u,u,u,26,27,24,25,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6,7,8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,2,2,3,10,9,11,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm26, %zmm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm0[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm8 = ymm1[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm8[0],ymm5[1],ymm8[2,3],ymm5[4],ymm8[5,6,7,8],ymm5[9],ymm8[10,11],ymm5[12],ymm8[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,23,26,27,u,u,24,25,26,27,u,u,26,27,26,27]
; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,2,2,3,8,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm2 ^ (zmm23 & (zmm1 ^ zmm2))
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4],ymm1[5],ymm0[6,7,8,9],ymm1[10],ymm0[11,12],ymm1[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm20 = [2,2,2,3,8,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm2 ^ (zmm28 & (zmm1 ^ zmm2))
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [5,0,0,0,6,0,0,6]
-; AVX512DQ-FCP-NEXT: vpermd %ymm5, %ymm1, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm3 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT: vpbroadcastd 72(%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vpandn %ymm1, %ymm3, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-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-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,0,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm0 = [5,0,0,0,6,0,0,6]
+; AVX512DQ-FCP-NEXT: vpermd %ymm4, %ymm0, %ymm0
; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm9[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5],ymm1[6],ymm2[7,8,9,10],ymm1[11],ymm2[12,13],ymm1[14],ymm2[15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm9[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm16, %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm14[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa %ymm11, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm15, %ymm3
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm14[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0,1],ymm4[2],ymm3[3,4],ymm4[5],ymm3[6,7,8,9],ymm4[10],ymm3[11,12],ymm4[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm2 ^ (zmm23 & (zmm0 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm8[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vprold $16, %ymm10, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm8[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm11 = [2,1,3,2,10,10,10,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm16
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512DQ-FCP-NEXT: vpermd %zmm16, %zmm0, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm15
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm15, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm21
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm28[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm28[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-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-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa %ymm7, %ymm10
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm29[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-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-FCP-NEXT: vpunpcklwd {{.*#+}} ymm3 = ymm0[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm7
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm29[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = 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-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm1 & (zmm4 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm13
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm25, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm25
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm26[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4],ymm2[5],ymm3[6,7,8,9],ymm2[10],ymm3[11,12],ymm2[13],ymm3[14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm13, %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm13, %ymm18
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm26[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = 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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [2,2,3,3,10,9,11,10]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm27, %zmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
+; AVX512DQ-FCP-NEXT: vpbroadcastd 72(%rax), %ymm0
+; AVX512DQ-FCP-NEXT: vpandn %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rax), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm0, %zmm2, %zmm0
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm12 & (zmm0 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm4))
+; AVX512DQ-FCP-NEXT: vpbroadcastd 8(%rax), %ymm0
+; AVX512DQ-FCP-NEXT: vpandn %ymm0, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm7, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm0 = ymm29[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = 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-FCP-NEXT: vmovdqa 64(%rcx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %xmm4
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm12 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, %xmm5
+; AVX512DQ-FCP-NEXT: vpbroadcastd {{.*#+}} ymm1 = [18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm18
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm29
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-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-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,0,3,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm4
; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2,3],xmm2[4],xmm4[5,6],xmm2[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [8,9,9,0,0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm24, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, (%rsp) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm5, %xmm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm1, %xmm1
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,0,0,1,8,8,9,0]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %xmm6
; AVX512DQ-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm6, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm3[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2],xmm5[3,4],xmm4[5],xmm5[6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm26 = [0,0,1,1,8,8,0,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm26, %zmm5
-; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (zmm4 & (zmm5 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm6
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm7 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm0 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,0,0,1,8,9,9,0]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm29, %zmm2
-; AVX512DQ-FCP-NEXT: vpbroadcastd 64(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vpbroadcastd 68(%rax), %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm30
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm20 & (zmm30 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm2 & (zmm30 ^ zmm5))
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm7, (%rsp) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm3 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm6, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm5 = xmm7[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm1[1],xmm5[2,3],xmm1[4],xmm5[5,6],xmm1[7]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,0,1,1,8,8,0,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm5
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm0 ^ (zmm24 & (zmm5 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm6
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm6
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,3,3,0,0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermq %zmm6, %zmm14, %zmm6
+; AVX512DQ-FCP-NEXT: vpbroadcastd 100(%rax), %ymm8
+; AVX512DQ-FCP-NEXT: vpbroadcastd 104(%rax), %ymm9
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm19 & (zmm0 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm17 & (zmm0 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %xmm8
; AVX512DQ-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm3, %xmm5
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm8[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[1,1,2,2]
; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3],xmm5[4],xmm6[5,6],xmm5[7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm24, %zmm5
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm12
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [8,9,9,0,0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm8, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm12
; AVX512DQ-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT: vprold $16, %xmm12, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm8 = xmm3[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm8[0,1],xmm6[2],xmm8[3,4],xmm6[5],xmm8[6,7]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm8, %xmm8
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm26, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm24
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm4 & (zmm8 ^ zmm5))
-; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm5
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm6 = xmm9[0,1],xmm6[2],xmm9[3,4],xmm6[5],xmm9[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm27
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm7, %zmm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm5 ^ (zmm6 & (zmm9 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r9), %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 64(%r8), %xmm11
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm11[4],xmm5[4],xmm11[5],xmm5[5],xmm11[6],xmm5[6],xmm11[7],xmm5[7]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm5[0],xmm11[1],xmm5[1],xmm11[2],xmm5[2],xmm11[3],xmm5[3]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm5
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm12 = [0,0,0,1,2,3,3,0]
+; AVX512DQ-FCP-NEXT: vpermq %zmm5, %zmm12, %zmm5
+; AVX512DQ-FCP-NEXT: vpbroadcastd 64(%rax), %ymm11
+; AVX512DQ-FCP-NEXT: vpbroadcastd 68(%rax), %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm11, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm21 & (zmm0 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm5 & (zmm0 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm11, %xmm9
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm13[0],xmm9[1],xmm13[2,3],xmm9[4],xmm13[5,6],xmm9[7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm8, %zmm9
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm1, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm13 = xmm0[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm8 = xmm13[0,1],xmm8[2],xmm13[3,4],xmm8[5],xmm13[6,7]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm13 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm13, %xmm13
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm7, %zmm13
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm9 ^ (zmm6 & (zmm13 ^ zmm9))
+; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm8
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm6, %ymm6
+; AVX512DQ-FCP-NEXT: vpermq %zmm6, %zmm12, %zmm6
+; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm8
+; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm9
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm21 & (zmm1 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm5 & (zmm1 ^ zmm13))
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX512DQ-FCP-NEXT: vprold $16, %xmm2, %xmm5
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,3]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm4, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm10
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm10, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm4 = xmm8[1,1,2,2]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2,3],xmm2[4],xmm4[5,6],xmm2[7]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm8[4],xmm10[4],xmm8[5],xmm10[5],xmm8[6],xmm10[6],xmm8[7],xmm10[7]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm7, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm5 ^ (zmm24 & (zmm2 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm4
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm5
-; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm4, %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm7, %xmm26
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm29, %zmm4
-; AVX512DQ-FCP-NEXT: vpbroadcastd (%rax), %ymm5
-; AVX512DQ-FCP-NEXT: vpbroadcastd 4(%rax), %ymm6
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm29
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (zmm20 & (zmm29 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm29 = zmm29 ^ (zmm2 & (zmm29 ^ zmm8))
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm21 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm21[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6,7,8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14,15]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm4
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm21[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-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-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm20[2,2,2,2,6,6,6,6]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm2[0,1],ymm5[2],ymm2[3,4],ymm5[5],ymm2[6,7,8,9],ymm5[10],ymm2[11,12],ymm5[13],ymm2[14,15]
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm2 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm20[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-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-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm2
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rcx), %xmm12
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdx), %xmm11
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,2,2,3,8,9,9,0]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm9, %zmm13
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm5, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm28[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm6[0,1,2],ymm5[3],ymm6[4,5],ymm5[6],ymm6[7,8,9,10],ymm5[11],ymm6[12,13],ymm5[14],ymm6[15]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm4 ^ (zmm1 & (zmm2 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vprold $16, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm17 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm17[1,2,2,3,5,6,6,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1],ymm1[2],ymm4[3,4],ymm1[5],ymm4[6,7,8,9],ymm1[10],ymm4[11,12],ymm1[13],ymm4[14,15]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,18,19,20,21,18,19,20,21]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm5, %ymm28
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[0,0,2,1,4,4,6,5]
-; AVX512DQ-FCP-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-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rsi), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%rdi), %xmm5
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm19 = [2,1,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm19, %zmm8
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm13 ^ (zmm23 & (zmm8 ^ zmm13))
-; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm22 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
-; AVX512DQ-FCP-NEXT: # ymm22 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm4[4],xmm1[4],xmm4[5],xmm1[5],xmm4[6],xmm1[6],xmm4[7],xmm1[7]
+; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vpermq %zmm1, %zmm14, %zmm0
+; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm1
+; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm30
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm19 & (zmm30 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm30 = zmm30 ^ (zmm17 & (zmm30 ^ zmm2))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 %ymm22, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = 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-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm15[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5],ymm0[6],ymm1[7,8,9,10],ymm0[11],ymm1[12,13],ymm0[14],ymm1[15]
+; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm6 = [151522058,0,421010202,421010202]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm15[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3],ymm1[4],ymm2[5,6,7,8],ymm1[9],ymm2[10,11],ymm1[12],ymm2[13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm3
+; AVX512DQ-FCP-NEXT: vmovdqu64 %ymm25, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm22 = [218894094,0,488382238,488382238]
; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm6
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm3, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 %ymm23, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm2 = ymm23[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27,22,23,26,27,0,0,24,25,26,27,0,0,26,27,26,27]
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm23[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1],ymm4[2],ymm2[3,4],ymm4[5],ymm2[6,7,8,9],ymm4[10],ymm2[11,12],ymm4[13],ymm2[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm1 ^ (zmm28 & (zmm2 ^ zmm1))
+; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm2
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-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-FCP-NEXT: vprold $16, %ymm18, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4],ymm1[5],ymm3[6,7,8,9],ymm1[10],ymm3[11,12],ymm1[13],ymm3[14,15]
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [2,1,3,2,10,10,10,11]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm17
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512DQ-FCP-NEXT: vpermd %zmm17, %zmm0, %zmm21
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm21 ^ (mem & (zmm21 ^ zmm1))
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm19
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm24 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm24[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm26
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm24[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm4[2],ymm1[3,4],ymm4[5],ymm1[6,7,8,9],ymm4[10],ymm1[11,12],ymm4[13],ymm1[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm25
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm4 = ymm12[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1],ymm4[2],ymm0[3,4],ymm4[5],ymm0[6,7,8,9],ymm4[10],ymm0[11,12],ymm4[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm4 = ymm11[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm7 = ymm12[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm4[2],ymm7[3,4],ymm4[5],ymm7[6,7,8,9],ymm4[10],ymm7[11,12],ymm4[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm1 ^ (zmm18 & (zmm7 ^ zmm1))
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm16 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa64 %ymm29, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm1 = ymm16[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = 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-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm0
; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm6 = ymm25[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm0[0,1],ymm6[2],ymm0[3,4],ymm6[5],ymm0[6,7,8,9],ymm6[10],ymm0[11,12],ymm6[13],ymm0[14,15]
-; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
-; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpermd %zmm16, %zmm0, %zmm16
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (mem & (zmm16 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm0
-; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm0, %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm3[14,15,12,13,u,u,u,u,u,u,u,u,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm20[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm2[1],ymm3[2,3],ymm2[4],ymm3[5,6,7,8],ymm2[9],ymm3[10,11],ymm2[12],ymm3[13,14,15]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm3 = ymm21[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm3[0,1,2],ymm2[3],ymm3[4,5],ymm2[6],ymm3[7,8,9,10],ymm2[11],ymm3[12,13],ymm2[14],ymm3[15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm10
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm9
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm15, %xmm27
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm14
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [2,2,2,3,8,8,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r9), %xmm4
-; AVX512DQ-FCP-NEXT: vmovdqa 96(%r8), %xmm3
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm13
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm15
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm0 ^ (zmm23 & (zmm14 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm18 = [6,7,3,3,7,7,6,7]
-; AVX512DQ-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm18, %ymm0 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpbroadcastd 96(%rax), %ymm19
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm0, %zmm19
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm23 & (zmm19 ^ zmm6))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm26 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm19 = zmm19 ^ (zmm26 & (zmm19 ^ zmm8))
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm1, %xmm0
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm5[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm1
-; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm5
-; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm1
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm5 = ymm17[3,3,3,3,7,7,7,7]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0,1],ymm5[2],ymm1[3,4],ymm5[5],ymm1[6,7,8,9],ymm5[10],ymm1[11,12],ymm5[13],ymm1[14,15]
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm8
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm6
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm5, %xmm15
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm21 = [0,0,0,1,8,8,9,0]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm0
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm20 = [6,7,4,5,0,0,8,9,6,7,4,5,0,0,8,9]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm15
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm12, %xmm12
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm12 = xmm11[0],xmm12[1],xmm11[2,3],xmm12[4],xmm11[5,6],xmm12[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, %zmm25
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm24, %zmm12
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm28 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm0 ^ (zmm28 & (zmm12 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm0 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm22 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm22 = ymm22 ^ (ymm23 & (ymm22 ^ ymm2))
-; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm2, %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} xmm24 = [0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm3
-; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm7, %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm15
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm13[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm15[0],ymm0[1],ymm15[2,3],ymm0[4],ymm15[5,6,7,8],ymm0[9],ymm15[10,11],ymm0[12],ymm15[13,14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm15
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm31
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-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]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm20, %zmm13
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm20 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm20[2,2,2,2,6,6,6,6]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm0[0,1],ymm15[2],ymm0[3,4],ymm15[5],ymm0[6,7,8,9],ymm15[10],ymm0[11,12],ymm15[13],ymm0[14,15]
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm0 = ymm6[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm20[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0,1],ymm0[2],ymm14[3,4],ymm0[5],ymm14[6,7,8,9],ymm0[10],ymm14[11,12],ymm0[13],ymm14[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm9, %zmm0
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,2,3,3,10,9,11,10]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm14, %zmm1
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm13 ^ (zmm18 & (zmm0 ^ zmm13))
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vprold $16, %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %ymm27 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm13 = ymm27[1,2,2,3,5,6,6,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0,1],ymm4[2],ymm13[3,4],ymm4[5],ymm13[6,7,8,9],ymm4[10],ymm13[11,12],ymm4[13],ymm13[14,15]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm9, %ymm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm9, %ymm25
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm27[0,0,2,1,4,4,6,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0,1,2],ymm15[3],ymm4[4,5],ymm15[6],ymm4[7,8,9,10],ymm15[11],ymm4[12,13],ymm15[14],ymm4[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm14, %zmm4
+; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm13
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm14 = [0,0,4,0,0,0,5,0,0,5,0,0,0,6,0,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm13, %zmm14, %zmm18
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (zmm28 & (zmm18 ^ zmm1))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm18 ^ (zmm1 & (zmm18 ^ zmm7))
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsp), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm7 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm7 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm9
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm11, %ymm14
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm15 = ymm12[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm14 = ymm15[0],ymm14[1],ymm15[2,3],ymm14[4],ymm15[5,6,7,8],ymm14[9],ymm15[10,11],ymm14[12],ymm15[13,14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,2,3,0,1,2,3,6,7,4,5,6,7,4,5]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm22 = [0,2,2,3,8,9,9,0]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm22, %zmm14
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm7 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm7 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm12 = ymm24[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1,2],ymm15[3],ymm12[4,5],ymm15[6],ymm12[7,8,9,10],ymm15[11],ymm12[12,13],ymm15[14],ymm12[15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,4,5,4,5,6,7,10,11,8,9,10,11,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [2,1,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm12
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm23 = [65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm14 ^ (zmm23 & (zmm12 ^ zmm14))
+; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31,26,27,28,29,26,27,28,29,26,27,28,29,30,31,30,31]
+; AVX512DQ-FCP-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm3
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm7
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm14 = ymm16[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0,1],ymm14[2],ymm7[3,4],ymm14[5],ymm7[6,7,8,9],ymm14[10],ymm7[11,12],ymm14[13],ymm7[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,2,3,6,7,4,5,6,7,4,5,12,13,14,15,0,1,2,3,8,9,10,11,14,15,12,13,14,15,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm14
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [2,2,2,3,8,8,8,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm7
+; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [0,5,4,0,0,6,5,0,0,5,4,0,0,6,5,0]
+; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpermd %zmm17, %zmm14, %zmm16
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (zmm28 & (zmm16 ^ zmm4))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm16 = zmm16 ^ (zmm1 & (zmm16 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm2 = [6,7,3,3,7,7,6,7]
+; AVX512DQ-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm1 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpbroadcastd 96(%rax), %ymm4
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm1, %zmm17
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (zmm4 & (zmm17 ^ zmm7))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm19 = [0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm17 = zmm17 ^ (zmm19 & (zmm17 ^ zmm12))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (ymm4 & (ymm3 ^ ymm0))
+; AVX512DQ-FCP-NEXT: vextracti64x4 $1, %zmm0, %ymm7
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} ymm10 = ymm0[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[3,3,3,3]
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = 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-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,2]
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm14
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm11
+; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm20[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3],ymm0[4],ymm9[5,6,7,8],ymm0[9],ymm9[10,11],ymm0[12],ymm9[13,14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm8 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm8 = xmm5[0],mem[0],xmm5[1],mem[1],xmm5[2],mem[2],xmm5[3],mem[3]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm31, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,14,15,12,13,u,u,u,u,30,31,28,29,u,u,u,u,30,31,28,29]
+; AVX512DQ-FCP-NEXT: vpshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm6 = mem[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm6[0,1,2],ymm9[3],ymm6[4,5],ymm9[6],ymm6[7,8,9,10],ymm9[11],ymm6[12,13],ymm9[14],ymm6[15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm6
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm0 ^ (zmm23 & (zmm6 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm0, %xmm0
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm27 = [8,9,9,0,0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm0
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufhw {{.*#+}} ymm3 = ymm4[0,1,2,3,5,5,7,6,8,9,10,11,13,13,15,14]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[3,3,3,3]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,24,25,22,23,22,23,u,u,u,u,u,u,u,u]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,2]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm7[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm7, %xmm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm15, %xmm31
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm17 = ymm7[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm18, %ymm7 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm15
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm23 & (zmm7 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vpbroadcastd 100(%rax), %ymm1
-; AVX512DQ-FCP-NEXT: vpbroadcastd 104(%rax), %ymm15
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm1, %zmm15
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm18 & (zmm15 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 ^ (zmm26 & (zmm7 ^ zmm14))
-; AVX512DQ-FCP-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-FCP-NEXT: vpternlogq {{.*#+}} zmm15 = zmm15 ^ (zmm0 & (zmm15 ^ zmm12))
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vprold $16, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[1,1,2,3]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0,1],xmm9[2],xmm10[3,4],xmm9[5],xmm10[6,7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm9
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm10, %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm10 = xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
-; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[1,1,2,2]
-; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm1 = xmm12[0],xmm1[1],xmm12[2,3],xmm1[4],xmm12[5,6],xmm1[7]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm9 ^ (zmm28 & (zmm1 ^ zmm9))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm24, %xmm9
-; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm5, %xmm5
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{.*#+}} xmm6 = xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm8
-; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm6
-; AVX512DQ-FCP-NEXT: vpbroadcastd 36(%rax), %ymm5
-; AVX512DQ-FCP-NEXT: vpbroadcastd 40(%rax), %ymm8
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm18 & (zmm5 ^ zmm6))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm0 & (zmm5 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogd {{.*#+}} ymm2 = m32bcst ^ (ymm0 & (ymm2 ^ m32bcst))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm0 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm3 = ymm3 ^ (ymm0 & (ymm3 ^ ymm2))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm5
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm8
+; AVX512DQ-FCP-NEXT: vpshufd {{.*#+}} ymm9 = ymm27[3,3,3,3,7,7,7,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4],ymm9[5],ymm8[6,7,8,9],ymm9[10],ymm8[11,12],ymm9[13],ymm8[14,15]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm29, %zmm8
+; AVX512DQ-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpbroadcastd 32(%rax), %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm8))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm19 & (zmm0 ^ zmm6))
+; AVX512DQ-FCP-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-FCP-NEXT: vpternlogd {{.*#+}} ymm7 = m32bcst ^ (ymm1 & (ymm7 ^ m32bcst))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm1 = [0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm10 = ymm10 ^ (ymm1 & (ymm10 ^ ymm7))
; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535,65535,65535,65535,65535,65535,0,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (ymm6 & (ymm4 ^ ymm22))
-; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm4[0,1,2,3],zmm3[0,1,2,3]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
-; AVX512DQ-FCP-NEXT: vpermd {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm1 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm1 = (zmm1 & zmm6) | mem
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm17, %zmm3 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm3 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm3 = (zmm3 & zmm6) | mem
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd (%rsp), %xmm6, %xmm6 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm12 = ymm12 ^ (ymm6 & (ymm12 ^ ymm3))
+; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm12[0,1,2,3],zmm10[0,1,2,3]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,13,0,0,0,14,0,0,14,0,0,0,15,0,0,15]
+; AVX512DQ-FCP-NEXT: vpermd %zmm13, %zmm3, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm2))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm14, %zmm2 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm2 = (zmm2 & zmm6) | mem
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm7
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm6 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm6 = (zmm6 & zmm7) | mem
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload
; AVX512DQ-FCP-NEXT: # xmm8 = xmm8[4],mem[4],xmm8[5],mem[5],xmm8[6],mem[6],xmm8[7],mem[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # xmm9 = xmm9[4],mem[4],xmm9[5],mem[5],xmm9[6],mem[6],xmm9[7],mem[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm6, %xmm6
-; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm9, %xmm9
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm10 = ymm10[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,1,3,4,5,5,7]
-; AVX512DQ-FCP-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]
-; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
-; AVX512DQ-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # ymm12 = mem[1,1,1,1,5,5,5,5]
-; AVX512DQ-FCP-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-FCP-NEXT: vmovdqa 32(%rax), %ymm12
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [0,1,4,5,4,5,5,7]
-; AVX512DQ-FCP-NEXT: vpermd %ymm12, %ymm13, %ymm13
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm12[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm12[16,17],zero,zero
-; AVX512DQ-FCP-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-FCP-NEXT: vpandn %ymm13, %ymm14, %ymm13
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm12
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,2,3,3,4,5,6,7]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,2,1]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,1,3]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,1,3,2]
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm13 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm14 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm13))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6, %zmm6 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm4 ^ (zmm28 & (zmm6 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm4 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm8 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm4 ^ (zmm28 & (zmm8 ^ zmm4))
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm4 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm1 ^ (zmm4 & (zmm6 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm3 ^ (zmm4 & (zmm8 ^ zmm3))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm1, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm9 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm10 # 16-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # xmm10 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [4,5,2,3,4,5,6,7,8,9,10,11,10,11,8,9]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm8, %xmm8
+; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm10, %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} ymm11 = ymm3[0,0,1,1,2,2,3,3,8,8,9,9,10,10,11,11]
+; AVX512DQ-FCP-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,1,3,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm12[0,1],ymm11[2],ymm12[3,4],ymm11[5],ymm12[6,7,8,9],ymm11[10],ymm12[11,12],ymm11[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,20,21,18,19,20,21,24,25,26,27,22,23,22,23]
+; AVX512DQ-FCP-NEXT: vpshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # ymm13 = mem[1,1,1,1,5,5,5,5]
+; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm12 = ymm12[0,1],ymm13[2],ymm12[3,4],ymm13[5],ymm12[6,7,8,9],ymm13[10],ymm12[11,12],ymm13[13],ymm12[14,15]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %ymm13
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm14 = [0,1,4,5,4,5,5,7]
+; AVX512DQ-FCP-NEXT: vpermd %ymm13, %ymm14, %ymm14
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm13[12,13],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,ymm13[16,17],zero,zero
+; AVX512DQ-FCP-NEXT: vpandnq %ymm14, %ymm28, %ymm14
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm13
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[0,2,3,3,4,5,6,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,2,1]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,1,3]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,1,3,2]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm3, %zmm3
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm1 ^ (mem & (zmm3 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm1 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm1 = zmm1 | (zmm0 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0, %zmm12 # 64-byte Folded Reload
-; AVX512DQ-FCP-NEXT: # zmm12 = zmm12 | (zmm0 & mem)
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 ^ (mem & (zmm12 ^ zmm3))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm21 = zmm21 ^ (mem & (zmm21 ^ zmm3))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8, %zmm7 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm14 = [65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535,65535,0,0,65535,65535,65535,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm5 ^ (zmm14 & (zmm7 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9, %zmm5 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm8 # 32-byte Folded Reload
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm14 & (zmm8 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535,65535,65535,65535,0,0,0,65535]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm2 ^ (zmm5 & (zmm7 ^ zmm2))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm6 ^ (zmm5 & (zmm8 ^ zmm6))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm5, %zmm5
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm2 ^ (mem & (zmm5 ^ zmm2))
+; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm2 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm2 = zmm2 | (zmm1 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq $248, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm13 # 64-byte Folded Reload
+; AVX512DQ-FCP-NEXT: # zmm13 = zmm13 | (zmm1 & mem)
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm13 ^ (mem & (zmm13 ^ zmm5))
; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 320(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 256(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 192(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 320(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 128(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, (%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 448(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 704(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 640(%rax)
-; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
-; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 576(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 384(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 640(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 384(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 256(%rax)
; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 64(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 512(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 832(%rax)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 768(%rax)
-; AVX512DQ-FCP-NEXT: addq $1480, %rsp # imm = 0x5C8
+; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovaps %zmm0, (%rax)
+; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 448(%rax)
+; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
+; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 704(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 512(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 832(%rax)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
+; AVX512DQ-FCP-NEXT: addq $1768, %rsp # imm = 0x6E8
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-5.ll
index ad9db98711a62..9c0e1d07f7c51 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
@@ -4836,165 +4836,172 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512-FCP-LABEL: store_i8_stride5_vf64:
; AVX512-FCP: # %bb.0:
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm0
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm3, %ymm2
-; AVX512-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[u],zero,xmm1[u,10],zero,xmm1[12],zero,xmm1[u,11]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,xmm1[8,u],zero,xmm1[7],zero,xmm1[9,u,11,u],zero,xmm1[10],zero,xmm1[12,u],zero
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm30
-; AVX512-FCP-NEXT: vporq %xmm0, %xmm2, %xmm22
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm6
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm7
-; AVX512-FCP-NEXT: vporq %ymm6, %ymm7, %ymm23
-; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm1[6],zero,xmm1[8,u],zero,xmm1[7],zero,xmm1[9],zero,xmm1[11,u],zero,xmm1[10],zero,xmm1[12]
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm31
-; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm1[6],zero,xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[11],zero,xmm1[u,10],zero,xmm1[12],zero
-; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm21
-; AVX512-FCP-NEXT: vporq %xmm6, %xmm7, %xmm24
-; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm7
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm4
-; AVX512-FCP-NEXT: vporq %ymm7, %ymm4, %ymm25
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm12[8],zero,xmm12[u,7],zero,xmm12[9],zero,xmm12[u],zero,xmm12[u,10],zero,xmm12[12],zero,xmm12[u,11]
+; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm30
+; AVX512-FCP-NEXT: vporq %xmm0, %xmm1, %xmm22
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm13, %ymm1
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm0
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm2
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm2, %ymm23
+; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512-FCP-NEXT: vpshufb %xmm2, %xmm6, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm31
+; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm2
+; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512-FCP-NEXT: vporq %xmm1, %xmm2, %xmm24
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
; AVX512-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm4
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm9
-; AVX512-FCP-NEXT: vporq %ymm4, %ymm9, %ymm26
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm4
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm13
-; AVX512-FCP-NEXT: vporq %ymm4, %ymm13, %ymm27
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
-; AVX512-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm7
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
-; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm13
-; AVX512-FCP-NEXT: vporq %ymm7, %ymm13, %ymm28
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
-; AVX512-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm15
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm15, %ymm16
-; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm3
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
-; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm3
-; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm8
-; AVX512-FCP-NEXT: vporq %ymm3, %ymm8, %ymm19
-; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm3
-; AVX512-FCP-NEXT: vporq %ymm2, %ymm3, %ymm18
-; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm1
-; AVX512-FCP-NEXT: vporq %ymm0, %ymm1, %ymm20
-; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm0
-; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm2
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm25
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm11, %ymm2
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm11
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm11, %zmm26
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512-FCP-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm2
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm15
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm15, %zmm27
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm28
+; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm7
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm9
+; AVX512-FCP-NEXT: vporq %ymm7, %ymm9, %ymm16
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm13, %ymm7
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm7, %ymm17
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm1
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm9
+; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm9, %ymm15
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm15, %ymm18
+; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm3
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm3, %ymm19
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm1
+; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512-FCP-NEXT: vporq %ymm1, %ymm0, %ymm20
+; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm9, %ymm0
+; AVX512-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm1
; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[u],zero,xmm1[u,10],zero,xmm1[12],zero,xmm1[u,11]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm2[8,u],zero,xmm2[7],zero,xmm2[9,u,11,u],zero,xmm2[10],zero,xmm2[12,u],zero
-; AVX512-FCP-NEXT: vpor %xmm3, %xmm4, %xmm5
+; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[8],zero,xmm2[u,7],zero,xmm2[9],zero,xmm2[u],zero,xmm2[u,10],zero,xmm2[12],zero,xmm2[u,11]
+; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm3
+; AVX512-FCP-NEXT: vpor %xmm1, %xmm3, %xmm4
; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm3[6],zero,xmm3[8,u],zero,xmm3[7],zero,xmm3[9],zero,xmm3[11,u],zero,xmm3[10],zero,xmm3[12]
-; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm4[6],zero,xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[11],zero,xmm4[u,10],zero,xmm4[12],zero
-; AVX512-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm7
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [1,1,2,2,2,2,2,2]
-; AVX512-FCP-NEXT: vpermd %ymm7, %ymm9, %ymm9
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512-FCP-NEXT: vpandn %ymm9, %ymm10, %ymm9
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm7
-; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm9
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [4,0,5,5,5,5,0,6]
-; AVX512-FCP-NEXT: vpermd %ymm9, %ymm11, %ymm11
-; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm8
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512-FCP-NEXT: vpandn %ymm11, %ymm9, %ymm11
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
-; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm11
-; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm12
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3],xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,2,3,3,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm20
-; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm11
-; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm14
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3],xmm11[4],xmm14[4],xmm11[5],xmm14[5],xmm11[6],xmm14[6],xmm11[7],xmm14[7]
-; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
-; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm0
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm1
-; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
-; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,0,1,1,8,8,9,9]
-; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm3, %zmm1
-; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm3, %zmm2
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm1
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512-FCP-NEXT: vpermd %zmm1, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm7
+; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm1
+; AVX512-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
+; AVX512-FCP-NEXT: vmovdqa64 %xmm29, %xmm9
+; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm7, %xmm9
+; AVX512-FCP-NEXT: vpor %xmm1, %xmm9, %xmm9
+; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [1,1,2,2,2,2,2,2]
+; AVX512-FCP-NEXT: vpermd %ymm1, %ymm13, %ymm13
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm14 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512-FCP-NEXT: vpandn %ymm13, %ymm14, %ymm13
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm13, %zmm1
+; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm13
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [4,0,5,5,5,5,0,6]
+; AVX512-FCP-NEXT: vpermd %ymm13, %ymm15, %ymm15
+; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm11
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512-FCP-NEXT: vpandn %ymm15, %ymm13, %ymm15
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3],xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [2,2,3,3,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm15, %zmm20
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
+; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm6, %xmm6
+; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm15, %zmm0
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
+; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm2, %xmm2
+; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3],xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
+; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,1,1,8,8,9,9]
+; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm2
+; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm5, %zmm3
; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm2))
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm22[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm24[0,0,1,1]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm4, %zmm4
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm5 & (zmm4 ^ zmm2))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 | (zmm4 & zmm10)
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm26[2,2,3,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm25, %zmm2
-; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm28[2,2,3,3]
-; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm27, %zmm4
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm5 & (zmm4 ^ zmm2))
-; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,2,3,3,10,10,11,11]
-; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm17
-; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm18
-; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm17 ^ (zmm2 & (zmm18 ^ zmm17))
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
-; AVX512-FCP-NEXT: vpermd %zmm1, %zmm5, %zmm5
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm9 & (zmm5 ^ zmm18))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm20 ^ (zmm2 & (zmm0 ^ zmm20))
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm4 & mem)
-; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512-FCP-NEXT: vpermd %zmm1, %zmm2, %zmm1
-; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 256(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 64(%r9)
-; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 192(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm2
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512-FCP-NEXT: vpermd %zmm2, %zmm4, %zmm4
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm22[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm24[0,0,1,1]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm5, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm6 & (zmm5 ^ zmm3))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm5 & zmm14)
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vporq %zmm3, %zmm5, %zmm3
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm28[2,2,3,3,6,6,7,7]
+; AVX512-FCP-NEXT: vporq %zmm5, %zmm7, %zmm5
+; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm7 & (zmm5 ^ zmm3))
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
+; AVX512-FCP-NEXT: vpermd %zmm2, %zmm3, %zmm3
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm13 & (zmm3 ^ zmm5))
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm17[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm16, %zmm5
+; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm19[2,2,3,3]
+; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm18, %zmm8
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm6 & (zmm8 ^ zmm5))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm20 ^ (zmm7 & (zmm0 ^ zmm20))
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 | (zmm8 & mem)
+; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512-FCP-NEXT: vpermd %zmm2, %zmm5, %zmm2
+; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 64(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 256(%r9)
+; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%r9)
; AVX512-FCP-NEXT: vzeroupper
; AVX512-FCP-NEXT: retq
;
@@ -5178,165 +5185,172 @@ define void @store_i8_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
;
; AVX512DQ-FCP-LABEL: store_i8_stride5_vf64:
; AVX512DQ-FCP: # %bb.0:
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm8
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm8, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm3
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm3, %ymm2
-; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm2, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[u],zero,xmm1[u,10],zero,xmm1[12],zero,xmm1[u,11]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm29
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm2 = zero,xmm1[8,u],zero,xmm1[7],zero,xmm1[9,u,11,u],zero,xmm1[10],zero,xmm1[12,u],zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm30
-; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm2, %xmm22
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm2
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm5
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm7
-; AVX512DQ-FCP-NEXT: vporq %ymm6, %ymm7, %ymm23
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm1[6],zero,xmm1[8,u],zero,xmm1[7],zero,xmm1[9],zero,xmm1[11,u],zero,xmm1[10],zero,xmm1[12]
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm31
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm1
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm1[6],zero,xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[11],zero,xmm1[u,10],zero,xmm1[12],zero
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm21
-; AVX512DQ-FCP-NEXT: vporq %xmm6, %xmm7, %xmm24
-; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm6
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm1
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm4
-; AVX512DQ-FCP-NEXT: vporq %ymm7, %ymm4, %ymm25
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm1 = [128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm1, %ymm18
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm5
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [12,13,128,15,12,13,14,128,12,13,14,15,128,u,u,u,16,128,18,19,16,17,128,19,16,17,18,128,16,17,18,19]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm19
+; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm21
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm12
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm12[8],zero,xmm12[u,7],zero,xmm12[9],zero,xmm12[u],zero,xmm12[u,10],zero,xmm12[12],zero,xmm12[u,11]
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm10
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,8,u,128,7,128,9,u,11,u,128,10,128,12,u,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm10, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm30
+; AVX512DQ-FCP-NEXT: vporq %xmm0, %xmm1, %xmm22
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm13
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128,128]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm0
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [u,u,12,13,128,u,u,u,14,128,u,u,14,15,128,u,u,u,16,128,u,u,16,17,128,u,u,u,18,128,u,u]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm2, %ymm23
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm6
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [128,6,128,8,u,128,7,128,9,128,11,u,128,10,128,12]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm2, %xmm6, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm31
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm8
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [6,128,8,128,u,7,128,9,128,11,128,u,10,128,12,128]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm29
+; AVX512DQ-FCP-NEXT: vporq %xmm1, %xmm2, %xmm24
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
; AVX512DQ-FCP-NEXT: # ymm14 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm6, %ymm4
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
-; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm9
-; AVX512DQ-FCP-NEXT: vporq %ymm4, %ymm9, %ymm26
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm4
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm10
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm13
-; AVX512DQ-FCP-NEXT: vporq %ymm4, %ymm13, %ymm27
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
-; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm10, %ymm7
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
-; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm13
-; AVX512DQ-FCP-NEXT: vporq %ymm7, %ymm13, %ymm28
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128,128,128,26,128,28,128,128,27,128,29,128,31,128,128,30,128]
-; AVX512DQ-FCP-NEXT: # ymm13 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
-; AVX512DQ-FCP-NEXT: # ymm7 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm15
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm15, %ymm16
-; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm8, %ymm3
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm3, %ymm17
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
-; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm5, %ymm1
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128,128,21,128,128,20,128,22,128,24,128,128,23,128,25,128,128]
; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1]
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm8
-; AVX512DQ-FCP-NEXT: vporq %ymm3, %ymm8, %ymm19
-; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm2
-; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm3
-; AVX512DQ-FCP-NEXT: vporq %ymm2, %ymm3, %ymm18
-; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm1
-; AVX512DQ-FCP-NEXT: vporq %ymm0, %ymm1, %ymm20
-; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm0
-; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm5, %ymm2
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm25
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30,27,0,128,26,128,28,0,128,0,128,29,128,31,0,128,30]
+; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm11, %ymm2
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0,19,128,21,0,128,20,128,22,128,24,0,128,23,128,25,0]
+; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm11
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm11, %zmm26
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128,128,27,128,128,26,128,28,128,30,128,128,29,128,31,128,128]
+; AVX512DQ-FCP-NEXT: # ymm11 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm2
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128,128,128,19,128,21,128,128,20,128,22,128,24,128,128,23,128]
+; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm13, %ymm15
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm15, %zmm27
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0,25,128,27,0,128,26,128,28,128,30,0,128,29,128,31,0]
+; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25,18,19,128,21,128,21,20,128,22,128,24,128,22,23,128,25]
+; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1]
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm28
+; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm13, %ymm9
+; AVX512DQ-FCP-NEXT: vporq %ymm7, %ymm9, %ymm16
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm13, %ymm7
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm7, %ymm17
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm7, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm9
+; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm9, %ymm15
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm15, %ymm18
+; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm7, %ymm3
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm3, %ymm19
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm1
+; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0
+; AVX512DQ-FCP-NEXT: vporq %ymm1, %ymm0, %ymm20
+; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm9, %ymm0
+; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm1
; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1
-; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm2
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[8],zero,xmm1[u,7],zero,xmm1[9],zero,xmm1[u],zero,xmm1[u,10],zero,xmm1[12],zero,xmm1[u,11]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = zero,xmm2[8,u],zero,xmm2[7],zero,xmm2[9,u,11,u],zero,xmm2[10],zero,xmm2[12,u],zero
-; AVX512DQ-FCP-NEXT: vpor %xmm3, %xmm4, %xmm5
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm2
+; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm5
+; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[8],zero,xmm2[u,7],zero,xmm2[9],zero,xmm2[u],zero,xmm2[u,10],zero,xmm2[12],zero,xmm2[u,11]
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm3
+; AVX512DQ-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm3
+; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm3, %xmm4
; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm3
-; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm4
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = zero,xmm3[6],zero,xmm3[8,u],zero,xmm3[7],zero,xmm3[9],zero,xmm3[11,u],zero,xmm3[10],zero,xmm3[12]
-; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm7 = xmm4[6],zero,xmm4[8],zero,xmm4[u,7],zero,xmm4[9],zero,xmm4[11],zero,xmm4[u,10],zero,xmm4[12],zero
-; AVX512DQ-FCP-NEXT: vpor %xmm6, %xmm7, %xmm6
-; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm9 = [1,1,2,2,2,2,2,2]
-; AVX512DQ-FCP-NEXT: vpermd %ymm7, %ymm9, %ymm9
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm7
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
-; AVX512DQ-FCP-NEXT: vpandn %ymm9, %ymm10, %ymm9
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm9, %zmm7
-; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm9
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm11 = [4,0,5,5,5,5,0,6]
-; AVX512DQ-FCP-NEXT: vpermd %ymm9, %ymm11, %ymm11
-; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm9, %ymm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpandn %ymm11, %ymm9, %ymm11
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm8, %zmm8
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm12
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3],xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm12 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm11
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [2,2,3,3,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm20
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm11
-; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm30, %xmm14
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3],xmm11[4],xmm14[4],xmm11[5],xmm14[5],xmm11[6],xmm14[6],xmm11[7],xmm14[7]
-; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm11, %xmm11
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm0
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm1, %xmm1
-; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
-; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm2, %xmm2
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,0,1,1,8,8,9,9]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm3, %zmm1
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm3, %zmm2
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm1))
-; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm1
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
-; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm7
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm1
+; AVX512DQ-FCP-NEXT: vpshufb %xmm1, %xmm3, %xmm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm9
+; AVX512DQ-FCP-NEXT: vpshufb %xmm9, %xmm7, %xmm9
+; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm9, %xmm9
+; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [12,128,128,128,128,13,128,128,128,128,14,128,128,128,128,15,128,128,128,128,16,128,128,128,128,17,128,128,128,128,18,128]
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm13 = [1,1,2,2,2,2,2,2]
+; AVX512DQ-FCP-NEXT: vpermd %ymm1, %ymm13, %ymm13
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm14 = [255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255]
+; AVX512DQ-FCP-NEXT: vpandn %ymm13, %ymm14, %ymm13
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm13, %zmm1
+; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm13
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm15 = [4,0,5,5,5,5,0,6]
+; AVX512DQ-FCP-NEXT: vpermd %ymm13, %ymm15, %ymm15
+; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm13, %ymm11
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm13 = [255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0,255,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpandn %ymm15, %ymm13, %ymm15
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm11, %zmm11
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3],xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [2,u,1,0,5,4,u,3,u,7,6,11,10,u,9,8]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [2,2,3,3,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm15, %zmm20
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7]
+; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm10 = [0,1,4,5,u,2,3,6,7,10,11,u,8,9,12,13]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm6, %xmm6
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm15, %zmm0
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm2, %xmm2
+; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3],xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7]
+; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm3, %xmm3
+; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,1,1,8,8,9,9]
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm5, %zmm2
+; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm5, %zmm3
; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (mem & (zmm3 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm22[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm24[0,0,1,1]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm4, %zmm4
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm5 & (zmm4 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm7 | (zmm4 & zmm10)
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm26[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm25, %zmm2
-; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm28[2,2,3,3]
-; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm27, %zmm4
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm2 ^ (zmm5 & (zmm4 ^ zmm2))
-; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [2,2,3,3,10,10,11,11]
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm17
-; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm18
-; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm2 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm18 = zmm17 ^ (zmm2 & (zmm18 ^ zmm17))
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
-; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm5, %zmm5
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm9 & (zmm5 ^ zmm18))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm20 ^ (zmm2 & (zmm0 ^ zmm20))
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm4 & mem)
-; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm2 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
-; AVX512DQ-FCP-NEXT: vpermd %zmm1, %zmm2, %zmm1
-; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm0))
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 128(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 256(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 64(%r9)
-; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 192(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm2
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,1,1,1,1,0,2,2,2,2,0]
+; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm4, %zmm4
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (mem & (zmm4 ^ zmm3))
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm22[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm24[0,0,1,1]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm5, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm6 & (zmm5 ^ zmm3))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm5 & zmm14)
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm3 = zmm25[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm26[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm3, %zmm5, %zmm3
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm5 = zmm27[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm7 = zmm28[2,2,3,3,6,6,7,7]
+; AVX512DQ-FCP-NEXT: vporq %zmm5, %zmm7, %zmm5
+; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm7 = [0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255,0,0,255,255,255]
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm5 = zmm3 ^ (zmm7 & (zmm5 ^ zmm3))
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [12,0,13,13,13,13,0,14,14,14,14,0,15,15,15,15]
+; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm3, %zmm3
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm13 & (zmm3 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm17[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm16, %zmm5
+; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm19[2,2,3,3]
+; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm18, %zmm8
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm5 ^ (zmm6 & (zmm8 ^ zmm5))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm20 ^ (zmm7 & (zmm0 ^ zmm20))
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm11 | (zmm8 & mem)
+; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [6,6,6,0,7,7,7,7,0,8,8,8,8,0,9,9]
+; AVX512DQ-FCP-NEXT: vpermd %zmm2, %zmm5, %zmm2
+; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (mem & (zmm2 ^ zmm0))
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 128(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 64(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 256(%r9)
+; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 192(%r9)
; AVX512DQ-FCP-NEXT: vzeroupper
; AVX512DQ-FCP-NEXT: retq
;
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
index 09b97df9afead..2abc201705929 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll
@@ -4113,58 +4113,51 @@ define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512BW-NEXT: vmovdqa (%r11), %xmm0
-; AVX512BW-NEXT: vmovdqa 16(%r11), %xmm1
-; AVX512BW-NEXT: vmovdqa (%r10), %xmm2
-; AVX512BW-NEXT: vmovdqa 16(%r10), %xmm3
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512BW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512BW-NEXT: vmovdqa 16(%r9), %xmm6
-; AVX512BW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512BW-NEXT: vmovdqa 16(%r8), %xmm8
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
+; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512BW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512BW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512BW-NEXT: vmovdqa (%r9), %ymm3
+; AVX512BW-NEXT: vmovdqa (%r11), %xmm4
+; AVX512BW-NEXT: vmovdqa (%r10), %xmm5
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512BW-NEXT: vmovdqa (%r9), %xmm7
+; AVX512BW-NEXT: vmovdqa (%r8), %xmm8
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39]
-; AVX512BW-NEXT: vpermt2w %zmm4, %zmm10, %zmm9
-; AVX512BW-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512BW-NEXT: vmovdqa 16(%rcx), %xmm11
-; AVX512BW-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512BW-NEXT: vmovdqa 16(%rdx), %xmm13
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15]
-; AVX512BW-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512BW-NEXT: vmovdqa64 16(%rsi), %xmm16
-; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm17
-; AVX512BW-NEXT: vmovdqa64 16(%rdi), %xmm18
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15]
-; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
-; AVX512BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm19
-; AVX512BW-NEXT: movw $-21846, %cx # imm = 0xAAAA
-; AVX512BW-NEXT: kmovd %ecx, %k1
-; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512BW-NEXT: vpermt2w %zmm9, %zmm10, %zmm14
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
-; AVX512BW-NEXT: vpermt2w %zmm9, %zmm20, %zmm21
-; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1}
-; AVX512BW-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]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512BW-NEXT: vpermt2w %zmm1, %zmm10, %zmm3
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
-; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
-; AVX512BW-NEXT: vpermt2w %zmm1, %zmm20, %zmm6
-; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1}
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512BW-NEXT: vpermt2w %zmm0, %zmm10, %zmm1
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
-; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512BW-NEXT: vpermt2w %zmm0, %zmm20, %zmm2
-; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm21, 128(%rax)
-; AVX512BW-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512BW-NEXT: vpermt2w %zmm6, %zmm10, %zmm9
+; AVX512BW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512BW-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512BW-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512BW-NEXT: vmovdqa (%rdi), %xmm14
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
+; AVX512BW-NEXT: vpermt2w %zmm12, %zmm16, %zmm15
+; AVX512BW-NEXT: movw $-21846, %si # imm = 0xAAAA
+; AVX512BW-NEXT: kmovd %esi, %k1
+; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm15 {%k1}
+; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],mem[2,3]
+; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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]
+; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm12 = 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]
+; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm17 = [0,8,32,40,1,9,33,41,2,10,34,42,3,11,35,43,4,12,36,44,5,13,37,45,6,14,38,46,7,15,39,47]
+; AVX512BW-NEXT: vpermt2w %zmm9, %zmm17, %zmm12
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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]
+; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm0 = 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]
+; AVX512BW-NEXT: vpermt2w %zmm2, %zmm17, %zmm0
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512BW-NEXT: vpermt2w %zmm1, %zmm10, %zmm2
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
+; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512BW-NEXT: vpermt2w %zmm1, %zmm16, %zmm3
+; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
+; AVX512BW-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm12, 128(%rax)
+; AVX512BW-NEXT: vmovdqa64 %zmm15, 64(%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
;
@@ -4173,58 +4166,51 @@ define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm0
-; AVX512BW-FCP-NEXT: vmovdqa 16(%r11), %xmm1
-; AVX512BW-FCP-NEXT: vmovdqa (%r10), %xmm2
-; AVX512BW-FCP-NEXT: vmovdqa 16(%r10), %xmm3
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm5
-; AVX512BW-FCP-NEXT: vmovdqa 16(%r9), %xmm6
-; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm7
-; AVX512BW-FCP-NEXT: vmovdqa 16(%r8), %xmm8
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512BW-FCP-NEXT: vmovdqa (%r8), %ymm2
+; AVX512BW-FCP-NEXT: vmovdqa (%r9), %ymm3
+; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm4
+; AVX512BW-FCP-NEXT: vmovdqa (%r10), %xmm5
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm7
+; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm8
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm10, %zmm9
-; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512BW-FCP-NEXT: vmovdqa 16(%rcx), %xmm11
-; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512BW-FCP-NEXT: vmovdqa 16(%rdx), %xmm13
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15]
-; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm16
-; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm17
-; AVX512BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm18
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15]
-; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm14, %zmm20, %zmm19
-; AVX512BW-FCP-NEXT: movw $-21846, %cx # imm = 0xAAAA
-; AVX512BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1}
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm9, %zmm10, %zmm14
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm9, %zmm20, %zmm21
-; AVX512BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1}
-; AVX512BW-FCP-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]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm10, %zmm3
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
-; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm20, %zmm6
-; AVX512BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1}
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm0, %zmm10, %zmm1
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
-; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512BW-FCP-NEXT: vpermt2w %zmm0, %zmm20, %zmm2
-; AVX512BW-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax)
-; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512BW-FCP-NEXT: vpermt2w %zmm6, %zmm10, %zmm9
+; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm14
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm12, %zmm16, %zmm15
+; AVX512BW-FCP-NEXT: movw $-21846, %si # imm = 0xAAAA
+; AVX512BW-FCP-NEXT: kmovd %esi, %k1
+; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm15 {%k1}
+; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],mem[2,3]
+; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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]
+; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm12 = 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]
+; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm17 = [0,8,32,40,1,9,33,41,2,10,34,42,3,11,35,43,4,12,36,44,5,13,37,45,6,14,38,46,7,15,39,47]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm9, %zmm17, %zmm12
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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]
+; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm0 = 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]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm2, %zmm17, %zmm0
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm10, %zmm2
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
+; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm16, %zmm3
+; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 128(%rax)
+; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, 64(%rax)
; AVX512BW-FCP-NEXT: vzeroupper
; AVX512BW-FCP-NEXT: retq
;
@@ -4233,58 +4219,51 @@ define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm0
-; AVX512DQ-BW-NEXT: vmovdqa 16(%r11), %xmm1
-; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm2
-; AVX512DQ-BW-NEXT: vmovdqa 16(%r10), %xmm3
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm5
-; AVX512DQ-BW-NEXT: vmovdqa 16(%r9), %xmm6
-; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm7
-; AVX512DQ-BW-NEXT: vmovdqa 16(%r8), %xmm8
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %ymm3
+; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm4
+; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm5
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm7
+; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm8
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm4, %zmm10, %zmm9
-; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512DQ-BW-NEXT: vmovdqa 16(%rcx), %xmm11
-; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512DQ-BW-NEXT: vmovdqa 16(%rdx), %xmm13
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15]
-; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%rsi), %xmm16
-; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %xmm17
-; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdi), %xmm18
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15]
-; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm19
-; AVX512DQ-BW-NEXT: movw $-21846, %cx # imm = 0xAAAA
-; AVX512DQ-BW-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm9, %zmm10, %zmm14
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm9, %zmm20, %zmm21
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1}
-; AVX512DQ-BW-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-BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm10, %zmm3
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
-; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm20, %zmm6
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1}
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm0, %zmm10, %zmm1
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
-; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512DQ-BW-NEXT: vpermt2w %zmm0, %zmm20, %zmm2
-; AVX512DQ-BW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 128(%rax)
-; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512DQ-BW-NEXT: vpermt2w %zmm6, %zmm10, %zmm9
+; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm14
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
+; AVX512DQ-BW-NEXT: vpermt2w %zmm12, %zmm16, %zmm15
+; AVX512DQ-BW-NEXT: movw $-21846, %si # imm = 0xAAAA
+; AVX512DQ-BW-NEXT: kmovd %esi, %k1
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm15 {%k1}
+; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],mem[2,3]
+; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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]
+; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
+; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm17 = [0,8,32,40,1,9,33,41,2,10,34,42,3,11,35,43,4,12,36,44,5,13,37,45,6,14,38,46,7,15,39,47]
+; AVX512DQ-BW-NEXT: vpermt2w %zmm9, %zmm17, %zmm12
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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]
+; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
+; AVX512DQ-BW-NEXT: vpermt2w %zmm2, %zmm17, %zmm0
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm10, %zmm2
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
+; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm16, %zmm3
+; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 128(%rax)
+; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 64(%rax)
; AVX512DQ-BW-NEXT: vzeroupper
; AVX512DQ-BW-NEXT: retq
;
@@ -4293,58 +4272,51 @@ define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vec
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm0
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r11), %xmm1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r10), %xmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r10), %xmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm5
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r9), %xmm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm7
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r8), %xmm8
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %ymm2
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %ymm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm4
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r10), %xmm5
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm7
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm8
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm10, %zmm9
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm4
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rcx), %xmm11
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm12
-; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rdx), %xmm13
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15]
-; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm15
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm16
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm17
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm18
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15]
-; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm14, %zmm20, %zmm19
-; AVX512DQ-BW-FCP-NEXT: movw $-21846, %cx # imm = 0xAAAA
-; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1
-; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm9, %zmm10, %zmm14
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm9, %zmm20, %zmm21
-; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1}
-; AVX512DQ-BW-FCP-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-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm10, %zmm3
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15]
-; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm20, %zmm6
-; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm0, %zmm10, %zmm1
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
-; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7]
-; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm0, %zmm20, %zmm2
-; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1}
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax)
-; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%rax)
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm6, %zmm10, %zmm9
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm6
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm11
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15]
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm13
+; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm14
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm16 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm12, %zmm16, %zmm15
+; AVX512DQ-BW-FCP-NEXT: movw $-21846, %si # imm = 0xAAAA
+; AVX512DQ-BW-FCP-NEXT: kmovd %esi, %k1
+; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm15 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],mem[2,3]
+; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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]
+; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
+; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
+; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm17 = [0,8,32,40,1,9,33,41,2,10,34,42,3,11,35,43,4,12,36,44,5,13,37,45,6,14,38,46,7,15,39,47]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm9, %zmm17, %zmm12
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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]
+; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm2, %zmm17, %zmm0
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm10, %zmm2
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
+; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
+; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm16, %zmm3
+; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1}
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 128(%rax)
+; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, 64(%rax)
; AVX512DQ-BW-FCP-NEXT: vzeroupper
; AVX512DQ-BW-FCP-NEXT: retq
%in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
More information about the llvm-commits
mailing list